4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
6 * Swap reorganised 29.12.95, Stephen Tweedie.
7 * kswapd added: 7.1.96 sct
8 * Removed kswapd_ctl limits, and swap out as many pages as needed
9 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11 * Multiqueue VM started 5.8.00, Rik van Riel.
15 #include <linux/module.h>
16 #include <linux/gfp.h>
17 #include <linux/kernel_stat.h>
18 #include <linux/swap.h>
19 #include <linux/pagemap.h>
20 #include <linux/init.h>
21 #include <linux/highmem.h>
22 #include <linux/vmstat.h>
23 #include <linux/file.h>
24 #include <linux/writeback.h>
25 #include <linux/blkdev.h>
26 #include <linux/buffer_head.h> /* for try_to_release_page(),
27 buffer_heads_over_limit */
28 #include <linux/mm_inline.h>
29 #include <linux/backing-dev.h>
30 #include <linux/rmap.h>
31 #include <linux/topology.h>
32 #include <linux/cpu.h>
33 #include <linux/cpuset.h>
34 #include <linux/compaction.h>
35 #include <linux/notifier.h>
36 #include <linux/rwsem.h>
37 #include <linux/delay.h>
38 #include <linux/kthread.h>
39 #include <linux/freezer.h>
40 #include <linux/memcontrol.h>
41 #include <linux/delayacct.h>
42 #include <linux/sysctl.h>
43 #include <linux/oom.h>
44 #include <linux/prefetch.h>
46 #include <asm/tlbflush.h>
47 #include <asm/div64.h>
49 #include <linux/swapops.h>
53 #define CREATE_TRACE_POINTS
54 #include <trace/events/vmscan.h>
57 /* Incremented by the number of inactive pages that were scanned */
58 unsigned long nr_scanned
;
60 /* Number of pages freed so far during a call to shrink_zones() */
61 unsigned long nr_reclaimed
;
63 /* How many pages shrink_list() should reclaim */
64 unsigned long nr_to_reclaim
;
66 unsigned long hibernation_mode
;
68 /* This context's GFP mask */
73 /* Can mapped pages be reclaimed? */
76 /* Can pages be swapped as part of reclaim? */
81 /* Scan (total_size >> priority) pages at once */
85 * The memory cgroup that hit its limit and as a result is the
86 * primary target of this reclaim invocation.
88 struct mem_cgroup
*target_mem_cgroup
;
91 * Nodemask of nodes allowed by the caller. If NULL, all nodes
97 struct mem_cgroup_zone
{
98 struct mem_cgroup
*mem_cgroup
;
102 #define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
104 #ifdef ARCH_HAS_PREFETCH
105 #define prefetch_prev_lru_page(_page, _base, _field) \
107 if ((_page)->lru.prev != _base) { \
110 prev = lru_to_page(&(_page->lru)); \
111 prefetch(&prev->_field); \
115 #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
118 #ifdef ARCH_HAS_PREFETCHW
119 #define prefetchw_prev_lru_page(_page, _base, _field) \
121 if ((_page)->lru.prev != _base) { \
124 prev = lru_to_page(&(_page->lru)); \
125 prefetchw(&prev->_field); \
129 #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
133 * From 0 .. 100. Higher means more swappy.
135 int vm_swappiness
= 60;
136 long vm_total_pages
; /* The total number of pages which the VM controls */
138 static LIST_HEAD(shrinker_list
);
139 static DECLARE_RWSEM(shrinker_rwsem
);
141 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
142 static bool global_reclaim(struct scan_control
*sc
)
144 return !sc
->target_mem_cgroup
;
147 static bool global_reclaim(struct scan_control
*sc
)
153 static struct zone_reclaim_stat
*get_reclaim_stat(struct mem_cgroup_zone
*mz
)
155 return &mem_cgroup_zone_lruvec(mz
->zone
, mz
->mem_cgroup
)->reclaim_stat
;
158 static unsigned long zone_nr_lru_pages(struct mem_cgroup_zone
*mz
,
161 if (!mem_cgroup_disabled())
162 return mem_cgroup_zone_nr_lru_pages(mz
->mem_cgroup
,
163 zone_to_nid(mz
->zone
),
167 return zone_page_state(mz
->zone
, NR_LRU_BASE
+ lru
);
172 * Add a shrinker callback to be called from the vm
174 void register_shrinker(struct shrinker
*shrinker
)
176 atomic_long_set(&shrinker
->nr_in_batch
, 0);
177 down_write(&shrinker_rwsem
);
178 list_add_tail(&shrinker
->list
, &shrinker_list
);
179 up_write(&shrinker_rwsem
);
181 EXPORT_SYMBOL(register_shrinker
);
186 void unregister_shrinker(struct shrinker
*shrinker
)
188 down_write(&shrinker_rwsem
);
189 list_del(&shrinker
->list
);
190 up_write(&shrinker_rwsem
);
192 EXPORT_SYMBOL(unregister_shrinker
);
194 static inline int do_shrinker_shrink(struct shrinker
*shrinker
,
195 struct shrink_control
*sc
,
196 unsigned long nr_to_scan
)
198 sc
->nr_to_scan
= nr_to_scan
;
199 return (*shrinker
->shrink
)(shrinker
, sc
);
202 #define SHRINK_BATCH 128
204 * Call the shrink functions to age shrinkable caches
206 * Here we assume it costs one seek to replace a lru page and that it also
207 * takes a seek to recreate a cache object. With this in mind we age equal
208 * percentages of the lru and ageable caches. This should balance the seeks
209 * generated by these structures.
211 * If the vm encountered mapped pages on the LRU it increase the pressure on
212 * slab to avoid swapping.
214 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
216 * `lru_pages' represents the number of on-LRU pages in all the zones which
217 * are eligible for the caller's allocation attempt. It is used for balancing
218 * slab reclaim versus page reclaim.
220 * Returns the number of slab objects which we shrunk.
222 unsigned long shrink_slab(struct shrink_control
*shrink
,
223 unsigned long nr_pages_scanned
,
224 unsigned long lru_pages
)
226 struct shrinker
*shrinker
;
227 unsigned long ret
= 0;
229 if (nr_pages_scanned
== 0)
230 nr_pages_scanned
= SWAP_CLUSTER_MAX
;
232 if (!down_read_trylock(&shrinker_rwsem
)) {
233 /* Assume we'll be able to shrink next time */
238 list_for_each_entry(shrinker
, &shrinker_list
, list
) {
239 unsigned long long delta
;
245 long batch_size
= shrinker
->batch
? shrinker
->batch
248 max_pass
= do_shrinker_shrink(shrinker
, shrink
, 0);
253 * copy the current shrinker scan count into a local variable
254 * and zero it so that other concurrent shrinker invocations
255 * don't also do this scanning work.
257 nr
= atomic_long_xchg(&shrinker
->nr_in_batch
, 0);
260 delta
= (4 * nr_pages_scanned
) / shrinker
->seeks
;
262 do_div(delta
, lru_pages
+ 1);
264 if (total_scan
< 0) {
265 printk(KERN_ERR
"shrink_slab: %pF negative objects to "
267 shrinker
->shrink
, total_scan
);
268 total_scan
= max_pass
;
272 * We need to avoid excessive windup on filesystem shrinkers
273 * due to large numbers of GFP_NOFS allocations causing the
274 * shrinkers to return -1 all the time. This results in a large
275 * nr being built up so when a shrink that can do some work
276 * comes along it empties the entire cache due to nr >>>
277 * max_pass. This is bad for sustaining a working set in
280 * Hence only allow the shrinker to scan the entire cache when
281 * a large delta change is calculated directly.
283 if (delta
< max_pass
/ 4)
284 total_scan
= min(total_scan
, max_pass
/ 2);
287 * Avoid risking looping forever due to too large nr value:
288 * never try to free more than twice the estimate number of
291 if (total_scan
> max_pass
* 2)
292 total_scan
= max_pass
* 2;
294 trace_mm_shrink_slab_start(shrinker
, shrink
, nr
,
295 nr_pages_scanned
, lru_pages
,
296 max_pass
, delta
, total_scan
);
298 while (total_scan
>= batch_size
) {
301 nr_before
= do_shrinker_shrink(shrinker
, shrink
, 0);
302 shrink_ret
= do_shrinker_shrink(shrinker
, shrink
,
304 if (shrink_ret
== -1)
306 if (shrink_ret
< nr_before
)
307 ret
+= nr_before
- shrink_ret
;
308 count_vm_events(SLABS_SCANNED
, batch_size
);
309 total_scan
-= batch_size
;
315 * move the unused scan count back into the shrinker in a
316 * manner that handles concurrent updates. If we exhausted the
317 * scan, there is no need to do an update.
320 new_nr
= atomic_long_add_return(total_scan
,
321 &shrinker
->nr_in_batch
);
323 new_nr
= atomic_long_read(&shrinker
->nr_in_batch
);
325 trace_mm_shrink_slab_end(shrinker
, shrink_ret
, nr
, new_nr
);
327 up_read(&shrinker_rwsem
);
333 static inline int is_page_cache_freeable(struct page
*page
)
336 * A freeable page cache page is referenced only by the caller
337 * that isolated the page, the page cache radix tree and
338 * optional buffer heads at page->private.
340 return page_count(page
) - page_has_private(page
) == 2;
343 static int may_write_to_queue(struct backing_dev_info
*bdi
,
344 struct scan_control
*sc
)
346 if (current
->flags
& PF_SWAPWRITE
)
348 if (!bdi_write_congested(bdi
))
350 if (bdi
== current
->backing_dev_info
)
356 * We detected a synchronous write error writing a page out. Probably
357 * -ENOSPC. We need to propagate that into the address_space for a subsequent
358 * fsync(), msync() or close().
360 * The tricky part is that after writepage we cannot touch the mapping: nothing
361 * prevents it from being freed up. But we have a ref on the page and once
362 * that page is locked, the mapping is pinned.
364 * We're allowed to run sleeping lock_page() here because we know the caller has
367 static void handle_write_error(struct address_space
*mapping
,
368 struct page
*page
, int error
)
371 if (page_mapping(page
) == mapping
)
372 mapping_set_error(mapping
, error
);
376 /* possible outcome of pageout() */
378 /* failed to write page out, page is locked */
380 /* move page to the active list, page is locked */
382 /* page has been sent to the disk successfully, page is unlocked */
384 /* page is clean and locked */
389 * pageout is called by shrink_page_list() for each dirty page.
390 * Calls ->writepage().
392 static pageout_t
pageout(struct page
*page
, struct address_space
*mapping
,
393 struct scan_control
*sc
)
396 * If the page is dirty, only perform writeback if that write
397 * will be non-blocking. To prevent this allocation from being
398 * stalled by pagecache activity. But note that there may be
399 * stalls if we need to run get_block(). We could test
400 * PagePrivate for that.
402 * If this process is currently in __generic_file_aio_write() against
403 * this page's queue, we can perform writeback even if that
406 * If the page is swapcache, write it back even if that would
407 * block, for some throttling. This happens by accident, because
408 * swap_backing_dev_info is bust: it doesn't reflect the
409 * congestion state of the swapdevs. Easy to fix, if needed.
411 if (!is_page_cache_freeable(page
))
415 * Some data journaling orphaned pages can have
416 * page->mapping == NULL while being dirty with clean buffers.
418 if (page_has_private(page
)) {
419 if (try_to_free_buffers(page
)) {
420 ClearPageDirty(page
);
421 printk("%s: orphaned page\n", __func__
);
427 if (mapping
->a_ops
->writepage
== NULL
)
428 return PAGE_ACTIVATE
;
429 if (!may_write_to_queue(mapping
->backing_dev_info
, sc
))
432 if (clear_page_dirty_for_io(page
)) {
434 struct writeback_control wbc
= {
435 .sync_mode
= WB_SYNC_NONE
,
436 .nr_to_write
= SWAP_CLUSTER_MAX
,
438 .range_end
= LLONG_MAX
,
442 SetPageReclaim(page
);
443 res
= mapping
->a_ops
->writepage(page
, &wbc
);
445 handle_write_error(mapping
, page
, res
);
446 if (res
== AOP_WRITEPAGE_ACTIVATE
) {
447 ClearPageReclaim(page
);
448 return PAGE_ACTIVATE
;
451 if (!PageWriteback(page
)) {
452 /* synchronous write or broken a_ops? */
453 ClearPageReclaim(page
);
455 trace_mm_vmscan_writepage(page
, trace_reclaim_flags(page
));
456 inc_zone_page_state(page
, NR_VMSCAN_WRITE
);
464 * Same as remove_mapping, but if the page is removed from the mapping, it
465 * gets returned with a refcount of 0.
467 static int __remove_mapping(struct address_space
*mapping
, struct page
*page
)
469 BUG_ON(!PageLocked(page
));
470 BUG_ON(mapping
!= page_mapping(page
));
472 spin_lock_irq(&mapping
->tree_lock
);
474 * The non racy check for a busy page.
476 * Must be careful with the order of the tests. When someone has
477 * a ref to the page, it may be possible that they dirty it then
478 * drop the reference. So if PageDirty is tested before page_count
479 * here, then the following race may occur:
481 * get_user_pages(&page);
482 * [user mapping goes away]
484 * !PageDirty(page) [good]
485 * SetPageDirty(page);
487 * !page_count(page) [good, discard it]
489 * [oops, our write_to data is lost]
491 * Reversing the order of the tests ensures such a situation cannot
492 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
493 * load is not satisfied before that of page->_count.
495 * Note that if SetPageDirty is always performed via set_page_dirty,
496 * and thus under tree_lock, then this ordering is not required.
498 if (!page_freeze_refs(page
, 2))
500 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
501 if (unlikely(PageDirty(page
))) {
502 page_unfreeze_refs(page
, 2);
506 if (PageSwapCache(page
)) {
507 swp_entry_t swap
= { .val
= page_private(page
) };
508 __delete_from_swap_cache(page
);
509 spin_unlock_irq(&mapping
->tree_lock
);
510 swapcache_free(swap
, page
);
512 void (*freepage
)(struct page
*);
514 freepage
= mapping
->a_ops
->freepage
;
516 __delete_from_page_cache(page
);
517 spin_unlock_irq(&mapping
->tree_lock
);
518 mem_cgroup_uncharge_cache_page(page
);
520 if (freepage
!= NULL
)
527 spin_unlock_irq(&mapping
->tree_lock
);
532 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
533 * someone else has a ref on the page, abort and return 0. If it was
534 * successfully detached, return 1. Assumes the caller has a single ref on
537 int remove_mapping(struct address_space
*mapping
, struct page
*page
)
539 if (__remove_mapping(mapping
, page
)) {
541 * Unfreezing the refcount with 1 rather than 2 effectively
542 * drops the pagecache ref for us without requiring another
545 page_unfreeze_refs(page
, 1);
552 * putback_lru_page - put previously isolated page onto appropriate LRU list
553 * @page: page to be put back to appropriate lru list
555 * Add previously isolated @page to appropriate LRU list.
556 * Page may still be unevictable for other reasons.
558 * lru_lock must not be held, interrupts must be enabled.
560 void putback_lru_page(struct page
*page
)
563 int active
= !!TestClearPageActive(page
);
564 int was_unevictable
= PageUnevictable(page
);
566 VM_BUG_ON(PageLRU(page
));
569 ClearPageUnevictable(page
);
571 if (page_evictable(page
, NULL
)) {
573 * For evictable pages, we can use the cache.
574 * In event of a race, worst case is we end up with an
575 * unevictable page on [in]active list.
576 * We know how to handle that.
578 lru
= active
+ page_lru_base_type(page
);
579 lru_cache_add_lru(page
, lru
);
582 * Put unevictable pages directly on zone's unevictable
585 lru
= LRU_UNEVICTABLE
;
586 add_page_to_unevictable_list(page
);
588 * When racing with an mlock or AS_UNEVICTABLE clearing
589 * (page is unlocked) make sure that if the other thread
590 * does not observe our setting of PG_lru and fails
591 * isolation/check_move_unevictable_pages,
592 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
593 * the page back to the evictable list.
595 * The other side is TestClearPageMlocked() or shmem_lock().
601 * page's status can change while we move it among lru. If an evictable
602 * page is on unevictable list, it never be freed. To avoid that,
603 * check after we added it to the list, again.
605 if (lru
== LRU_UNEVICTABLE
&& page_evictable(page
, NULL
)) {
606 if (!isolate_lru_page(page
)) {
610 /* This means someone else dropped this page from LRU
611 * So, it will be freed or putback to LRU again. There is
612 * nothing to do here.
616 if (was_unevictable
&& lru
!= LRU_UNEVICTABLE
)
617 count_vm_event(UNEVICTABLE_PGRESCUED
);
618 else if (!was_unevictable
&& lru
== LRU_UNEVICTABLE
)
619 count_vm_event(UNEVICTABLE_PGCULLED
);
621 put_page(page
); /* drop ref from isolate */
624 enum page_references
{
626 PAGEREF_RECLAIM_CLEAN
,
631 static enum page_references
page_check_references(struct page
*page
,
632 struct scan_control
*sc
)
634 int referenced_ptes
, referenced_page
;
635 unsigned long vm_flags
;
637 referenced_ptes
= page_referenced(page
, 1, sc
->target_mem_cgroup
,
639 referenced_page
= TestClearPageReferenced(page
);
642 * Mlock lost the isolation race with us. Let try_to_unmap()
643 * move the page to the unevictable list.
645 if (vm_flags
& VM_LOCKED
)
646 return PAGEREF_RECLAIM
;
648 if (referenced_ptes
) {
649 if (PageSwapBacked(page
))
650 return PAGEREF_ACTIVATE
;
652 * All mapped pages start out with page table
653 * references from the instantiating fault, so we need
654 * to look twice if a mapped file page is used more
657 * Mark it and spare it for another trip around the
658 * inactive list. Another page table reference will
659 * lead to its activation.
661 * Note: the mark is set for activated pages as well
662 * so that recently deactivated but used pages are
665 SetPageReferenced(page
);
667 if (referenced_page
|| referenced_ptes
> 1)
668 return PAGEREF_ACTIVATE
;
671 * Activate file-backed executable pages after first usage.
673 if (vm_flags
& VM_EXEC
)
674 return PAGEREF_ACTIVATE
;
679 /* Reclaim if clean, defer dirty pages to writeback */
680 if (referenced_page
&& !PageSwapBacked(page
))
681 return PAGEREF_RECLAIM_CLEAN
;
683 return PAGEREF_RECLAIM
;
687 * shrink_page_list() returns the number of reclaimed pages
689 static unsigned long shrink_page_list(struct list_head
*page_list
,
691 struct scan_control
*sc
,
692 unsigned long *ret_nr_dirty
,
693 unsigned long *ret_nr_writeback
)
695 LIST_HEAD(ret_pages
);
696 LIST_HEAD(free_pages
);
698 unsigned long nr_dirty
= 0;
699 unsigned long nr_congested
= 0;
700 unsigned long nr_reclaimed
= 0;
701 unsigned long nr_writeback
= 0;
705 while (!list_empty(page_list
)) {
706 enum page_references references
;
707 struct address_space
*mapping
;
713 page
= lru_to_page(page_list
);
714 list_del(&page
->lru
);
716 if (!trylock_page(page
))
719 VM_BUG_ON(PageActive(page
));
720 VM_BUG_ON(page_zone(page
) != zone
);
724 if (unlikely(!page_evictable(page
, NULL
)))
727 if (!sc
->may_unmap
&& page_mapped(page
))
730 /* Double the slab pressure for mapped and swapcache pages */
731 if (page_mapped(page
) || PageSwapCache(page
))
734 may_enter_fs
= (sc
->gfp_mask
& __GFP_FS
) ||
735 (PageSwapCache(page
) && (sc
->gfp_mask
& __GFP_IO
));
737 if (PageWriteback(page
)) {
743 references
= page_check_references(page
, sc
);
744 switch (references
) {
745 case PAGEREF_ACTIVATE
:
746 goto activate_locked
;
749 case PAGEREF_RECLAIM
:
750 case PAGEREF_RECLAIM_CLEAN
:
751 ; /* try to reclaim the page below */
755 * Anonymous process memory has backing store?
756 * Try to allocate it some swap space here.
758 if (PageAnon(page
) && !PageSwapCache(page
)) {
759 if (!(sc
->gfp_mask
& __GFP_IO
))
761 if (!add_to_swap(page
))
762 goto activate_locked
;
766 mapping
= page_mapping(page
);
769 * The page is mapped into the page tables of one or more
770 * processes. Try to unmap it here.
772 if (page_mapped(page
) && mapping
) {
773 switch (try_to_unmap(page
, TTU_UNMAP
)) {
775 goto activate_locked
;
781 ; /* try to free the page below */
785 if (PageDirty(page
)) {
789 * Only kswapd can writeback filesystem pages to
790 * avoid risk of stack overflow but do not writeback
791 * unless under significant pressure.
793 if (page_is_file_cache(page
) &&
794 (!current_is_kswapd() ||
795 sc
->priority
>= DEF_PRIORITY
- 2)) {
797 * Immediately reclaim when written back.
798 * Similar in principal to deactivate_page()
799 * except we already have the page isolated
800 * and know it's dirty
802 inc_zone_page_state(page
, NR_VMSCAN_IMMEDIATE
);
803 SetPageReclaim(page
);
808 if (references
== PAGEREF_RECLAIM_CLEAN
)
812 if (!sc
->may_writepage
)
815 /* Page is dirty, try to write it out here */
816 switch (pageout(page
, mapping
, sc
)) {
821 goto activate_locked
;
823 if (PageWriteback(page
))
829 * A synchronous write - probably a ramdisk. Go
830 * ahead and try to reclaim the page.
832 if (!trylock_page(page
))
834 if (PageDirty(page
) || PageWriteback(page
))
836 mapping
= page_mapping(page
);
838 ; /* try to free the page below */
843 * If the page has buffers, try to free the buffer mappings
844 * associated with this page. If we succeed we try to free
847 * We do this even if the page is PageDirty().
848 * try_to_release_page() does not perform I/O, but it is
849 * possible for a page to have PageDirty set, but it is actually
850 * clean (all its buffers are clean). This happens if the
851 * buffers were written out directly, with submit_bh(). ext3
852 * will do this, as well as the blockdev mapping.
853 * try_to_release_page() will discover that cleanness and will
854 * drop the buffers and mark the page clean - it can be freed.
856 * Rarely, pages can have buffers and no ->mapping. These are
857 * the pages which were not successfully invalidated in
858 * truncate_complete_page(). We try to drop those buffers here
859 * and if that worked, and the page is no longer mapped into
860 * process address space (page_count == 1) it can be freed.
861 * Otherwise, leave the page on the LRU so it is swappable.
863 if (page_has_private(page
)) {
864 if (!try_to_release_page(page
, sc
->gfp_mask
))
865 goto activate_locked
;
866 if (!mapping
&& page_count(page
) == 1) {
868 if (put_page_testzero(page
))
872 * rare race with speculative reference.
873 * the speculative reference will free
874 * this page shortly, so we may
875 * increment nr_reclaimed here (and
876 * leave it off the LRU).
884 if (!mapping
|| !__remove_mapping(mapping
, page
))
888 * At this point, we have no other references and there is
889 * no way to pick any more up (removed from LRU, removed
890 * from pagecache). Can use non-atomic bitops now (and
891 * we obviously don't have to worry about waking up a process
892 * waiting on the page lock, because there are no references.
894 __clear_page_locked(page
);
899 * Is there need to periodically free_page_list? It would
900 * appear not as the counts should be low
902 list_add(&page
->lru
, &free_pages
);
906 if (PageSwapCache(page
))
907 try_to_free_swap(page
);
909 putback_lru_page(page
);
913 /* Not a candidate for swapping, so reclaim swap space. */
914 if (PageSwapCache(page
) && vm_swap_full())
915 try_to_free_swap(page
);
916 VM_BUG_ON(PageActive(page
));
922 list_add(&page
->lru
, &ret_pages
);
923 VM_BUG_ON(PageLRU(page
) || PageUnevictable(page
));
927 * Tag a zone as congested if all the dirty pages encountered were
928 * backed by a congested BDI. In this case, reclaimers should just
929 * back off and wait for congestion to clear because further reclaim
930 * will encounter the same problem
932 if (nr_dirty
&& nr_dirty
== nr_congested
&& global_reclaim(sc
))
933 zone_set_flag(zone
, ZONE_CONGESTED
);
935 free_hot_cold_page_list(&free_pages
, 1);
937 list_splice(&ret_pages
, page_list
);
938 count_vm_events(PGACTIVATE
, pgactivate
);
939 *ret_nr_dirty
+= nr_dirty
;
940 *ret_nr_writeback
+= nr_writeback
;
945 * Attempt to remove the specified page from its LRU. Only take this page
946 * if it is of the appropriate PageActive status. Pages which are being
947 * freed elsewhere are also ignored.
949 * page: page to consider
950 * mode: one of the LRU isolation modes defined above
952 * returns 0 on success, -ve errno on failure.
954 int __isolate_lru_page(struct page
*page
, isolate_mode_t mode
)
958 /* Only take pages on the LRU. */
962 /* Do not give back unevictable pages for compaction */
963 if (PageUnevictable(page
))
969 * To minimise LRU disruption, the caller can indicate that it only
970 * wants to isolate pages it will be able to operate on without
971 * blocking - clean pages for the most part.
973 * ISOLATE_CLEAN means that only clean pages should be isolated. This
974 * is used by reclaim when it is cannot write to backing storage
976 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
977 * that it is possible to migrate without blocking
979 if (mode
& (ISOLATE_CLEAN
|ISOLATE_ASYNC_MIGRATE
)) {
980 /* All the caller can do on PageWriteback is block */
981 if (PageWriteback(page
))
984 if (PageDirty(page
)) {
985 struct address_space
*mapping
;
987 /* ISOLATE_CLEAN means only clean pages */
988 if (mode
& ISOLATE_CLEAN
)
992 * Only pages without mappings or that have a
993 * ->migratepage callback are possible to migrate
996 mapping
= page_mapping(page
);
997 if (mapping
&& !mapping
->a_ops
->migratepage
)
1002 if ((mode
& ISOLATE_UNMAPPED
) && page_mapped(page
))
1005 if (likely(get_page_unless_zero(page
))) {
1007 * Be careful not to clear PageLRU until after we're
1008 * sure the page is not being freed elsewhere -- the
1009 * page release code relies on it.
1019 * zone->lru_lock is heavily contended. Some of the functions that
1020 * shrink the lists perform better by taking out a batch of pages
1021 * and working on them outside the LRU lock.
1023 * For pagecache intensive workloads, this function is the hottest
1024 * spot in the kernel (apart from copy_*_user functions).
1026 * Appropriate locks must be held before calling this function.
1028 * @nr_to_scan: The number of pages to look through on the list.
1029 * @lruvec: The LRU vector to pull pages from.
1030 * @dst: The temp list to put pages on to.
1031 * @nr_scanned: The number of pages that were scanned.
1032 * @sc: The scan_control struct for this reclaim session
1033 * @mode: One of the LRU isolation modes
1034 * @lru: LRU list id for isolating
1036 * returns how many pages were moved onto *@dst.
1038 static unsigned long isolate_lru_pages(unsigned long nr_to_scan
,
1039 struct lruvec
*lruvec
, struct list_head
*dst
,
1040 unsigned long *nr_scanned
, struct scan_control
*sc
,
1041 isolate_mode_t mode
, enum lru_list lru
)
1043 struct list_head
*src
;
1044 unsigned long nr_taken
= 0;
1046 int file
= is_file_lru(lru
);
1048 src
= &lruvec
->lists
[lru
];
1050 for (scan
= 0; scan
< nr_to_scan
&& !list_empty(src
); scan
++) {
1053 page
= lru_to_page(src
);
1054 prefetchw_prev_lru_page(page
, src
, flags
);
1056 VM_BUG_ON(!PageLRU(page
));
1058 switch (__isolate_lru_page(page
, mode
)) {
1060 mem_cgroup_lru_del_list(page
, lru
);
1061 list_move(&page
->lru
, dst
);
1062 nr_taken
+= hpage_nr_pages(page
);
1066 /* else it is being freed elsewhere */
1067 list_move(&page
->lru
, src
);
1077 trace_mm_vmscan_lru_isolate(sc
->order
,
1085 * isolate_lru_page - tries to isolate a page from its LRU list
1086 * @page: page to isolate from its LRU list
1088 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1089 * vmstat statistic corresponding to whatever LRU list the page was on.
1091 * Returns 0 if the page was removed from an LRU list.
1092 * Returns -EBUSY if the page was not on an LRU list.
1094 * The returned page will have PageLRU() cleared. If it was found on
1095 * the active list, it will have PageActive set. If it was found on
1096 * the unevictable list, it will have the PageUnevictable bit set. That flag
1097 * may need to be cleared by the caller before letting the page go.
1099 * The vmstat statistic corresponding to the list on which the page was
1100 * found will be decremented.
1103 * (1) Must be called with an elevated refcount on the page. This is a
1104 * fundamentnal difference from isolate_lru_pages (which is called
1105 * without a stable reference).
1106 * (2) the lru_lock must not be held.
1107 * (3) interrupts must be enabled.
1109 int isolate_lru_page(struct page
*page
)
1113 VM_BUG_ON(!page_count(page
));
1115 if (PageLRU(page
)) {
1116 struct zone
*zone
= page_zone(page
);
1118 spin_lock_irq(&zone
->lru_lock
);
1119 if (PageLRU(page
)) {
1120 int lru
= page_lru(page
);
1125 del_page_from_lru_list(zone
, page
, lru
);
1127 spin_unlock_irq(&zone
->lru_lock
);
1133 * Are there way too many processes in the direct reclaim path already?
1135 static int too_many_isolated(struct zone
*zone
, int file
,
1136 struct scan_control
*sc
)
1138 unsigned long inactive
, isolated
;
1140 if (current_is_kswapd())
1143 if (!global_reclaim(sc
))
1147 inactive
= zone_page_state(zone
, NR_INACTIVE_FILE
);
1148 isolated
= zone_page_state(zone
, NR_ISOLATED_FILE
);
1150 inactive
= zone_page_state(zone
, NR_INACTIVE_ANON
);
1151 isolated
= zone_page_state(zone
, NR_ISOLATED_ANON
);
1154 return isolated
> inactive
;
1157 static noinline_for_stack
void
1158 putback_inactive_pages(struct mem_cgroup_zone
*mz
,
1159 struct list_head
*page_list
)
1161 struct zone_reclaim_stat
*reclaim_stat
= get_reclaim_stat(mz
);
1162 struct zone
*zone
= mz
->zone
;
1163 LIST_HEAD(pages_to_free
);
1166 * Put back any unfreeable pages.
1168 while (!list_empty(page_list
)) {
1169 struct page
*page
= lru_to_page(page_list
);
1172 VM_BUG_ON(PageLRU(page
));
1173 list_del(&page
->lru
);
1174 if (unlikely(!page_evictable(page
, NULL
))) {
1175 spin_unlock_irq(&zone
->lru_lock
);
1176 putback_lru_page(page
);
1177 spin_lock_irq(&zone
->lru_lock
);
1181 lru
= page_lru(page
);
1182 add_page_to_lru_list(zone
, page
, lru
);
1183 if (is_active_lru(lru
)) {
1184 int file
= is_file_lru(lru
);
1185 int numpages
= hpage_nr_pages(page
);
1186 reclaim_stat
->recent_rotated
[file
] += numpages
;
1188 if (put_page_testzero(page
)) {
1189 __ClearPageLRU(page
);
1190 __ClearPageActive(page
);
1191 del_page_from_lru_list(zone
, page
, lru
);
1193 if (unlikely(PageCompound(page
))) {
1194 spin_unlock_irq(&zone
->lru_lock
);
1195 (*get_compound_page_dtor(page
))(page
);
1196 spin_lock_irq(&zone
->lru_lock
);
1198 list_add(&page
->lru
, &pages_to_free
);
1203 * To save our caller's stack, now use input list for pages to free.
1205 list_splice(&pages_to_free
, page_list
);
1208 static noinline_for_stack
void
1209 update_isolated_counts(struct mem_cgroup_zone
*mz
,
1210 struct list_head
*page_list
,
1211 unsigned long *nr_anon
,
1212 unsigned long *nr_file
)
1214 struct zone
*zone
= mz
->zone
;
1215 unsigned int count
[NR_LRU_LISTS
] = { 0, };
1216 unsigned long nr_active
= 0;
1221 * Count pages and clear active flags
1223 list_for_each_entry(page
, page_list
, lru
) {
1224 int numpages
= hpage_nr_pages(page
);
1225 lru
= page_lru_base_type(page
);
1226 if (PageActive(page
)) {
1228 ClearPageActive(page
);
1229 nr_active
+= numpages
;
1231 count
[lru
] += numpages
;
1235 __count_vm_events(PGDEACTIVATE
, nr_active
);
1237 __mod_zone_page_state(zone
, NR_ACTIVE_FILE
,
1238 -count
[LRU_ACTIVE_FILE
]);
1239 __mod_zone_page_state(zone
, NR_INACTIVE_FILE
,
1240 -count
[LRU_INACTIVE_FILE
]);
1241 __mod_zone_page_state(zone
, NR_ACTIVE_ANON
,
1242 -count
[LRU_ACTIVE_ANON
]);
1243 __mod_zone_page_state(zone
, NR_INACTIVE_ANON
,
1244 -count
[LRU_INACTIVE_ANON
]);
1246 *nr_anon
= count
[LRU_ACTIVE_ANON
] + count
[LRU_INACTIVE_ANON
];
1247 *nr_file
= count
[LRU_ACTIVE_FILE
] + count
[LRU_INACTIVE_FILE
];
1249 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
, *nr_anon
);
1250 __mod_zone_page_state(zone
, NR_ISOLATED_FILE
, *nr_file
);
1255 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1256 * of reclaimed pages
1258 static noinline_for_stack
unsigned long
1259 shrink_inactive_list(unsigned long nr_to_scan
, struct mem_cgroup_zone
*mz
,
1260 struct scan_control
*sc
, enum lru_list lru
)
1262 LIST_HEAD(page_list
);
1263 unsigned long nr_scanned
;
1264 unsigned long nr_reclaimed
= 0;
1265 unsigned long nr_taken
;
1266 unsigned long nr_anon
;
1267 unsigned long nr_file
;
1268 unsigned long nr_dirty
= 0;
1269 unsigned long nr_writeback
= 0;
1270 isolate_mode_t isolate_mode
= 0;
1271 int file
= is_file_lru(lru
);
1272 struct zone
*zone
= mz
->zone
;
1273 struct zone_reclaim_stat
*reclaim_stat
= get_reclaim_stat(mz
);
1274 struct lruvec
*lruvec
= mem_cgroup_zone_lruvec(zone
, mz
->mem_cgroup
);
1276 while (unlikely(too_many_isolated(zone
, file
, sc
))) {
1277 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
1279 /* We are about to die and free our memory. Return now. */
1280 if (fatal_signal_pending(current
))
1281 return SWAP_CLUSTER_MAX
;
1287 isolate_mode
|= ISOLATE_UNMAPPED
;
1288 if (!sc
->may_writepage
)
1289 isolate_mode
|= ISOLATE_CLEAN
;
1291 spin_lock_irq(&zone
->lru_lock
);
1293 nr_taken
= isolate_lru_pages(nr_to_scan
, lruvec
, &page_list
,
1294 &nr_scanned
, sc
, isolate_mode
, lru
);
1295 if (global_reclaim(sc
)) {
1296 zone
->pages_scanned
+= nr_scanned
;
1297 if (current_is_kswapd())
1298 __count_zone_vm_events(PGSCAN_KSWAPD
, zone
,
1301 __count_zone_vm_events(PGSCAN_DIRECT
, zone
,
1304 spin_unlock_irq(&zone
->lru_lock
);
1309 update_isolated_counts(mz
, &page_list
, &nr_anon
, &nr_file
);
1311 nr_reclaimed
= shrink_page_list(&page_list
, zone
, sc
,
1312 &nr_dirty
, &nr_writeback
);
1314 spin_lock_irq(&zone
->lru_lock
);
1316 reclaim_stat
->recent_scanned
[0] += nr_anon
;
1317 reclaim_stat
->recent_scanned
[1] += nr_file
;
1319 if (global_reclaim(sc
)) {
1320 if (current_is_kswapd())
1321 __count_zone_vm_events(PGSTEAL_KSWAPD
, zone
,
1324 __count_zone_vm_events(PGSTEAL_DIRECT
, zone
,
1328 putback_inactive_pages(mz
, &page_list
);
1330 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
, -nr_anon
);
1331 __mod_zone_page_state(zone
, NR_ISOLATED_FILE
, -nr_file
);
1333 spin_unlock_irq(&zone
->lru_lock
);
1335 free_hot_cold_page_list(&page_list
, 1);
1338 * If reclaim is isolating dirty pages under writeback, it implies
1339 * that the long-lived page allocation rate is exceeding the page
1340 * laundering rate. Either the global limits are not being effective
1341 * at throttling processes due to the page distribution throughout
1342 * zones or there is heavy usage of a slow backing device. The
1343 * only option is to throttle from reclaim context which is not ideal
1344 * as there is no guarantee the dirtying process is throttled in the
1345 * same way balance_dirty_pages() manages.
1347 * This scales the number of dirty pages that must be under writeback
1348 * before throttling depending on priority. It is a simple backoff
1349 * function that has the most effect in the range DEF_PRIORITY to
1350 * DEF_PRIORITY-2 which is the priority reclaim is considered to be
1351 * in trouble and reclaim is considered to be in trouble.
1353 * DEF_PRIORITY 100% isolated pages must be PageWriteback to throttle
1354 * DEF_PRIORITY-1 50% must be PageWriteback
1355 * DEF_PRIORITY-2 25% must be PageWriteback, kswapd in trouble
1357 * DEF_PRIORITY-6 For SWAP_CLUSTER_MAX isolated pages, throttle if any
1358 * isolated page is PageWriteback
1360 if (nr_writeback
&& nr_writeback
>=
1361 (nr_taken
>> (DEF_PRIORITY
- sc
->priority
)))
1362 wait_iff_congested(zone
, BLK_RW_ASYNC
, HZ
/10);
1364 trace_mm_vmscan_lru_shrink_inactive(zone
->zone_pgdat
->node_id
,
1366 nr_scanned
, nr_reclaimed
,
1368 trace_shrink_flags(file
));
1369 return nr_reclaimed
;
1373 * This moves pages from the active list to the inactive list.
1375 * We move them the other way if the page is referenced by one or more
1376 * processes, from rmap.
1378 * If the pages are mostly unmapped, the processing is fast and it is
1379 * appropriate to hold zone->lru_lock across the whole operation. But if
1380 * the pages are mapped, the processing is slow (page_referenced()) so we
1381 * should drop zone->lru_lock around each page. It's impossible to balance
1382 * this, so instead we remove the pages from the LRU while processing them.
1383 * It is safe to rely on PG_active against the non-LRU pages in here because
1384 * nobody will play with that bit on a non-LRU page.
1386 * The downside is that we have to touch page->_count against each page.
1387 * But we had to alter page->flags anyway.
1390 static void move_active_pages_to_lru(struct zone
*zone
,
1391 struct list_head
*list
,
1392 struct list_head
*pages_to_free
,
1395 unsigned long pgmoved
= 0;
1398 while (!list_empty(list
)) {
1399 struct lruvec
*lruvec
;
1401 page
= lru_to_page(list
);
1403 VM_BUG_ON(PageLRU(page
));
1406 lruvec
= mem_cgroup_lru_add_list(zone
, page
, lru
);
1407 list_move(&page
->lru
, &lruvec
->lists
[lru
]);
1408 pgmoved
+= hpage_nr_pages(page
);
1410 if (put_page_testzero(page
)) {
1411 __ClearPageLRU(page
);
1412 __ClearPageActive(page
);
1413 del_page_from_lru_list(zone
, page
, lru
);
1415 if (unlikely(PageCompound(page
))) {
1416 spin_unlock_irq(&zone
->lru_lock
);
1417 (*get_compound_page_dtor(page
))(page
);
1418 spin_lock_irq(&zone
->lru_lock
);
1420 list_add(&page
->lru
, pages_to_free
);
1423 __mod_zone_page_state(zone
, NR_LRU_BASE
+ lru
, pgmoved
);
1424 if (!is_active_lru(lru
))
1425 __count_vm_events(PGDEACTIVATE
, pgmoved
);
1428 static void shrink_active_list(unsigned long nr_to_scan
,
1429 struct mem_cgroup_zone
*mz
,
1430 struct scan_control
*sc
,
1433 unsigned long nr_taken
;
1434 unsigned long nr_scanned
;
1435 unsigned long vm_flags
;
1436 LIST_HEAD(l_hold
); /* The pages which were snipped off */
1437 LIST_HEAD(l_active
);
1438 LIST_HEAD(l_inactive
);
1440 struct zone_reclaim_stat
*reclaim_stat
= get_reclaim_stat(mz
);
1441 unsigned long nr_rotated
= 0;
1442 isolate_mode_t isolate_mode
= 0;
1443 int file
= is_file_lru(lru
);
1444 struct zone
*zone
= mz
->zone
;
1445 struct lruvec
*lruvec
= mem_cgroup_zone_lruvec(zone
, mz
->mem_cgroup
);
1450 isolate_mode
|= ISOLATE_UNMAPPED
;
1451 if (!sc
->may_writepage
)
1452 isolate_mode
|= ISOLATE_CLEAN
;
1454 spin_lock_irq(&zone
->lru_lock
);
1456 nr_taken
= isolate_lru_pages(nr_to_scan
, lruvec
, &l_hold
,
1457 &nr_scanned
, sc
, isolate_mode
, lru
);
1458 if (global_reclaim(sc
))
1459 zone
->pages_scanned
+= nr_scanned
;
1461 reclaim_stat
->recent_scanned
[file
] += nr_taken
;
1463 __count_zone_vm_events(PGREFILL
, zone
, nr_scanned
);
1464 __mod_zone_page_state(zone
, NR_LRU_BASE
+ lru
, -nr_taken
);
1465 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
+ file
, nr_taken
);
1466 spin_unlock_irq(&zone
->lru_lock
);
1468 while (!list_empty(&l_hold
)) {
1470 page
= lru_to_page(&l_hold
);
1471 list_del(&page
->lru
);
1473 if (unlikely(!page_evictable(page
, NULL
))) {
1474 putback_lru_page(page
);
1478 if (unlikely(buffer_heads_over_limit
)) {
1479 if (page_has_private(page
) && trylock_page(page
)) {
1480 if (page_has_private(page
))
1481 try_to_release_page(page
, 0);
1486 if (page_referenced(page
, 0, sc
->target_mem_cgroup
,
1488 nr_rotated
+= hpage_nr_pages(page
);
1490 * Identify referenced, file-backed active pages and
1491 * give them one more trip around the active list. So
1492 * that executable code get better chances to stay in
1493 * memory under moderate memory pressure. Anon pages
1494 * are not likely to be evicted by use-once streaming
1495 * IO, plus JVM can create lots of anon VM_EXEC pages,
1496 * so we ignore them here.
1498 if ((vm_flags
& VM_EXEC
) && page_is_file_cache(page
)) {
1499 list_add(&page
->lru
, &l_active
);
1504 ClearPageActive(page
); /* we are de-activating */
1505 list_add(&page
->lru
, &l_inactive
);
1509 * Move pages back to the lru list.
1511 spin_lock_irq(&zone
->lru_lock
);
1513 * Count referenced pages from currently used mappings as rotated,
1514 * even though only some of them are actually re-activated. This
1515 * helps balance scan pressure between file and anonymous pages in
1518 reclaim_stat
->recent_rotated
[file
] += nr_rotated
;
1520 move_active_pages_to_lru(zone
, &l_active
, &l_hold
, lru
);
1521 move_active_pages_to_lru(zone
, &l_inactive
, &l_hold
, lru
- LRU_ACTIVE
);
1522 __mod_zone_page_state(zone
, NR_ISOLATED_ANON
+ file
, -nr_taken
);
1523 spin_unlock_irq(&zone
->lru_lock
);
1525 free_hot_cold_page_list(&l_hold
, 1);
1529 static int inactive_anon_is_low_global(struct zone
*zone
)
1531 unsigned long active
, inactive
;
1533 active
= zone_page_state(zone
, NR_ACTIVE_ANON
);
1534 inactive
= zone_page_state(zone
, NR_INACTIVE_ANON
);
1536 if (inactive
* zone
->inactive_ratio
< active
)
1543 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1544 * @zone: zone to check
1545 * @sc: scan control of this context
1547 * Returns true if the zone does not have enough inactive anon pages,
1548 * meaning some active anon pages need to be deactivated.
1550 static int inactive_anon_is_low(struct mem_cgroup_zone
*mz
)
1553 * If we don't have swap space, anonymous page deactivation
1556 if (!total_swap_pages
)
1559 if (!mem_cgroup_disabled())
1560 return mem_cgroup_inactive_anon_is_low(mz
->mem_cgroup
,
1563 return inactive_anon_is_low_global(mz
->zone
);
1566 static inline int inactive_anon_is_low(struct mem_cgroup_zone
*mz
)
1572 static int inactive_file_is_low_global(struct zone
*zone
)
1574 unsigned long active
, inactive
;
1576 active
= zone_page_state(zone
, NR_ACTIVE_FILE
);
1577 inactive
= zone_page_state(zone
, NR_INACTIVE_FILE
);
1579 return (active
> inactive
);
1583 * inactive_file_is_low - check if file pages need to be deactivated
1584 * @mz: memory cgroup and zone to check
1586 * When the system is doing streaming IO, memory pressure here
1587 * ensures that active file pages get deactivated, until more
1588 * than half of the file pages are on the inactive list.
1590 * Once we get to that situation, protect the system's working
1591 * set from being evicted by disabling active file page aging.
1593 * This uses a different ratio than the anonymous pages, because
1594 * the page cache uses a use-once replacement algorithm.
1596 static int inactive_file_is_low(struct mem_cgroup_zone
*mz
)
1598 if (!mem_cgroup_disabled())
1599 return mem_cgroup_inactive_file_is_low(mz
->mem_cgroup
,
1602 return inactive_file_is_low_global(mz
->zone
);
1605 static int inactive_list_is_low(struct mem_cgroup_zone
*mz
, int file
)
1608 return inactive_file_is_low(mz
);
1610 return inactive_anon_is_low(mz
);
1613 static unsigned long shrink_list(enum lru_list lru
, unsigned long nr_to_scan
,
1614 struct mem_cgroup_zone
*mz
,
1615 struct scan_control
*sc
)
1617 int file
= is_file_lru(lru
);
1619 if (is_active_lru(lru
)) {
1620 if (inactive_list_is_low(mz
, file
))
1621 shrink_active_list(nr_to_scan
, mz
, sc
, lru
);
1625 return shrink_inactive_list(nr_to_scan
, mz
, sc
, lru
);
1628 static int vmscan_swappiness(struct scan_control
*sc
)
1630 if (global_reclaim(sc
))
1631 return vm_swappiness
;
1632 return mem_cgroup_swappiness(sc
->target_mem_cgroup
);
1636 * Determine how aggressively the anon and file LRU lists should be
1637 * scanned. The relative value of each set of LRU lists is determined
1638 * by looking at the fraction of the pages scanned we did rotate back
1639 * onto the active list instead of evict.
1641 * nr[0] = anon pages to scan; nr[1] = file pages to scan
1643 static void get_scan_count(struct mem_cgroup_zone
*mz
, struct scan_control
*sc
,
1646 unsigned long anon
, file
, free
;
1647 unsigned long anon_prio
, file_prio
;
1648 unsigned long ap
, fp
;
1649 struct zone_reclaim_stat
*reclaim_stat
= get_reclaim_stat(mz
);
1650 u64 fraction
[2], denominator
;
1653 bool force_scan
= false;
1656 * If the zone or memcg is small, nr[l] can be 0. This
1657 * results in no scanning on this priority and a potential
1658 * priority drop. Global direct reclaim can go to the next
1659 * zone and tends to have no problems. Global kswapd is for
1660 * zone balancing and it needs to scan a minimum amount. When
1661 * reclaiming for a memcg, a priority drop can cause high
1662 * latencies, so it's better to scan a minimum amount there as
1665 if (current_is_kswapd() && mz
->zone
->all_unreclaimable
)
1667 if (!global_reclaim(sc
))
1670 /* If we have no swap space, do not bother scanning anon pages. */
1671 if (!sc
->may_swap
|| (nr_swap_pages
<= 0)) {
1679 anon
= zone_nr_lru_pages(mz
, LRU_ACTIVE_ANON
) +
1680 zone_nr_lru_pages(mz
, LRU_INACTIVE_ANON
);
1681 file
= zone_nr_lru_pages(mz
, LRU_ACTIVE_FILE
) +
1682 zone_nr_lru_pages(mz
, LRU_INACTIVE_FILE
);
1684 if (global_reclaim(sc
)) {
1685 free
= zone_page_state(mz
->zone
, NR_FREE_PAGES
);
1686 /* If we have very few page cache pages,
1687 force-scan anon pages. */
1688 if (unlikely(file
+ free
<= high_wmark_pages(mz
->zone
))) {
1697 * With swappiness at 100, anonymous and file have the same priority.
1698 * This scanning priority is essentially the inverse of IO cost.
1700 anon_prio
= vmscan_swappiness(sc
);
1701 file_prio
= 200 - vmscan_swappiness(sc
);
1704 * OK, so we have swap space and a fair amount of page cache
1705 * pages. We use the recently rotated / recently scanned
1706 * ratios to determine how valuable each cache is.
1708 * Because workloads change over time (and to avoid overflow)
1709 * we keep these statistics as a floating average, which ends
1710 * up weighing recent references more than old ones.
1712 * anon in [0], file in [1]
1714 spin_lock_irq(&mz
->zone
->lru_lock
);
1715 if (unlikely(reclaim_stat
->recent_scanned
[0] > anon
/ 4)) {
1716 reclaim_stat
->recent_scanned
[0] /= 2;
1717 reclaim_stat
->recent_rotated
[0] /= 2;
1720 if (unlikely(reclaim_stat
->recent_scanned
[1] > file
/ 4)) {
1721 reclaim_stat
->recent_scanned
[1] /= 2;
1722 reclaim_stat
->recent_rotated
[1] /= 2;
1726 * The amount of pressure on anon vs file pages is inversely
1727 * proportional to the fraction of recently scanned pages on
1728 * each list that were recently referenced and in active use.
1730 ap
= anon_prio
* (reclaim_stat
->recent_scanned
[0] + 1);
1731 ap
/= reclaim_stat
->recent_rotated
[0] + 1;
1733 fp
= file_prio
* (reclaim_stat
->recent_scanned
[1] + 1);
1734 fp
/= reclaim_stat
->recent_rotated
[1] + 1;
1735 spin_unlock_irq(&mz
->zone
->lru_lock
);
1739 denominator
= ap
+ fp
+ 1;
1741 for_each_evictable_lru(lru
) {
1742 int file
= is_file_lru(lru
);
1745 scan
= zone_nr_lru_pages(mz
, lru
);
1746 if (sc
->priority
|| noswap
|| !vmscan_swappiness(sc
)) {
1747 scan
>>= sc
->priority
;
1748 if (!scan
&& force_scan
)
1749 scan
= SWAP_CLUSTER_MAX
;
1750 scan
= div64_u64(scan
* fraction
[file
], denominator
);
1756 /* Use reclaim/compaction for costly allocs or under memory pressure */
1757 static bool in_reclaim_compaction(struct scan_control
*sc
)
1759 if (COMPACTION_BUILD
&& sc
->order
&&
1760 (sc
->order
> PAGE_ALLOC_COSTLY_ORDER
||
1761 sc
->priority
< DEF_PRIORITY
- 2))
1768 * Reclaim/compaction is used for high-order allocation requests. It reclaims
1769 * order-0 pages before compacting the zone. should_continue_reclaim() returns
1770 * true if more pages should be reclaimed such that when the page allocator
1771 * calls try_to_compact_zone() that it will have enough free pages to succeed.
1772 * It will give up earlier than that if there is difficulty reclaiming pages.
1774 static inline bool should_continue_reclaim(struct mem_cgroup_zone
*mz
,
1775 unsigned long nr_reclaimed
,
1776 unsigned long nr_scanned
,
1777 struct scan_control
*sc
)
1779 unsigned long pages_for_compaction
;
1780 unsigned long inactive_lru_pages
;
1782 /* If not in reclaim/compaction mode, stop */
1783 if (!in_reclaim_compaction(sc
))
1786 /* Consider stopping depending on scan and reclaim activity */
1787 if (sc
->gfp_mask
& __GFP_REPEAT
) {
1789 * For __GFP_REPEAT allocations, stop reclaiming if the
1790 * full LRU list has been scanned and we are still failing
1791 * to reclaim pages. This full LRU scan is potentially
1792 * expensive but a __GFP_REPEAT caller really wants to succeed
1794 if (!nr_reclaimed
&& !nr_scanned
)
1798 * For non-__GFP_REPEAT allocations which can presumably
1799 * fail without consequence, stop if we failed to reclaim
1800 * any pages from the last SWAP_CLUSTER_MAX number of
1801 * pages that were scanned. This will return to the
1802 * caller faster at the risk reclaim/compaction and
1803 * the resulting allocation attempt fails
1810 * If we have not reclaimed enough pages for compaction and the
1811 * inactive lists are large enough, continue reclaiming
1813 pages_for_compaction
= (2UL << sc
->order
);
1814 inactive_lru_pages
= zone_nr_lru_pages(mz
, LRU_INACTIVE_FILE
);
1815 if (nr_swap_pages
> 0)
1816 inactive_lru_pages
+= zone_nr_lru_pages(mz
, LRU_INACTIVE_ANON
);
1817 if (sc
->nr_reclaimed
< pages_for_compaction
&&
1818 inactive_lru_pages
> pages_for_compaction
)
1821 /* If compaction would go ahead or the allocation would succeed, stop */
1822 switch (compaction_suitable(mz
->zone
, sc
->order
)) {
1823 case COMPACT_PARTIAL
:
1824 case COMPACT_CONTINUE
:
1832 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1834 static void shrink_mem_cgroup_zone(struct mem_cgroup_zone
*mz
,
1835 struct scan_control
*sc
)
1837 unsigned long nr
[NR_LRU_LISTS
];
1838 unsigned long nr_to_scan
;
1840 unsigned long nr_reclaimed
, nr_scanned
;
1841 unsigned long nr_to_reclaim
= sc
->nr_to_reclaim
;
1842 struct blk_plug plug
;
1846 nr_scanned
= sc
->nr_scanned
;
1847 get_scan_count(mz
, sc
, nr
);
1849 blk_start_plug(&plug
);
1850 while (nr
[LRU_INACTIVE_ANON
] || nr
[LRU_ACTIVE_FILE
] ||
1851 nr
[LRU_INACTIVE_FILE
]) {
1852 for_each_evictable_lru(lru
) {
1854 nr_to_scan
= min_t(unsigned long,
1855 nr
[lru
], SWAP_CLUSTER_MAX
);
1856 nr
[lru
] -= nr_to_scan
;
1858 nr_reclaimed
+= shrink_list(lru
, nr_to_scan
,
1863 * On large memory systems, scan >> priority can become
1864 * really large. This is fine for the starting priority;
1865 * we want to put equal scanning pressure on each zone.
1866 * However, if the VM has a harder time of freeing pages,
1867 * with multiple processes reclaiming pages, the total
1868 * freeing target can get unreasonably large.
1870 if (nr_reclaimed
>= nr_to_reclaim
&&
1871 sc
->priority
< DEF_PRIORITY
)
1874 blk_finish_plug(&plug
);
1875 sc
->nr_reclaimed
+= nr_reclaimed
;
1878 * Even if we did not try to evict anon pages at all, we want to
1879 * rebalance the anon lru active/inactive ratio.
1881 if (inactive_anon_is_low(mz
))
1882 shrink_active_list(SWAP_CLUSTER_MAX
, mz
,
1883 sc
, LRU_ACTIVE_ANON
);
1885 /* reclaim/compaction might need reclaim to continue */
1886 if (should_continue_reclaim(mz
, nr_reclaimed
,
1887 sc
->nr_scanned
- nr_scanned
, sc
))
1890 throttle_vm_writeout(sc
->gfp_mask
);
1893 static void shrink_zone(struct zone
*zone
, struct scan_control
*sc
)
1895 struct mem_cgroup
*root
= sc
->target_mem_cgroup
;
1896 struct mem_cgroup_reclaim_cookie reclaim
= {
1898 .priority
= sc
->priority
,
1900 struct mem_cgroup
*memcg
;
1902 memcg
= mem_cgroup_iter(root
, NULL
, &reclaim
);
1904 struct mem_cgroup_zone mz
= {
1905 .mem_cgroup
= memcg
,
1909 shrink_mem_cgroup_zone(&mz
, sc
);
1911 * Limit reclaim has historically picked one memcg and
1912 * scanned it with decreasing priority levels until
1913 * nr_to_reclaim had been reclaimed. This priority
1914 * cycle is thus over after a single memcg.
1916 * Direct reclaim and kswapd, on the other hand, have
1917 * to scan all memory cgroups to fulfill the overall
1918 * scan target for the zone.
1920 if (!global_reclaim(sc
)) {
1921 mem_cgroup_iter_break(root
, memcg
);
1924 memcg
= mem_cgroup_iter(root
, memcg
, &reclaim
);
1928 /* Returns true if compaction should go ahead for a high-order request */
1929 static inline bool compaction_ready(struct zone
*zone
, struct scan_control
*sc
)
1931 unsigned long balance_gap
, watermark
;
1934 /* Do not consider compaction for orders reclaim is meant to satisfy */
1935 if (sc
->order
<= PAGE_ALLOC_COSTLY_ORDER
)
1939 * Compaction takes time to run and there are potentially other
1940 * callers using the pages just freed. Continue reclaiming until
1941 * there is a buffer of free pages available to give compaction
1942 * a reasonable chance of completing and allocating the page
1944 balance_gap
= min(low_wmark_pages(zone
),
1945 (zone
->present_pages
+ KSWAPD_ZONE_BALANCE_GAP_RATIO
-1) /
1946 KSWAPD_ZONE_BALANCE_GAP_RATIO
);
1947 watermark
= high_wmark_pages(zone
) + balance_gap
+ (2UL << sc
->order
);
1948 watermark_ok
= zone_watermark_ok_safe(zone
, 0, watermark
, 0, 0);
1951 * If compaction is deferred, reclaim up to a point where
1952 * compaction will have a chance of success when re-enabled
1954 if (compaction_deferred(zone
, sc
->order
))
1955 return watermark_ok
;
1957 /* If compaction is not ready to start, keep reclaiming */
1958 if (!compaction_suitable(zone
, sc
->order
))
1961 return watermark_ok
;
1965 * This is the direct reclaim path, for page-allocating processes. We only
1966 * try to reclaim pages from zones which will satisfy the caller's allocation
1969 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
1971 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1973 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
1974 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
1975 * zone defense algorithm.
1977 * If a zone is deemed to be full of pinned pages then just give it a light
1978 * scan then give up on it.
1980 * This function returns true if a zone is being reclaimed for a costly
1981 * high-order allocation and compaction is ready to begin. This indicates to
1982 * the caller that it should consider retrying the allocation instead of
1985 static bool shrink_zones(struct zonelist
*zonelist
, struct scan_control
*sc
)
1989 unsigned long nr_soft_reclaimed
;
1990 unsigned long nr_soft_scanned
;
1991 bool aborted_reclaim
= false;
1994 * If the number of buffer_heads in the machine exceeds the maximum
1995 * allowed level, force direct reclaim to scan the highmem zone as
1996 * highmem pages could be pinning lowmem pages storing buffer_heads
1998 if (buffer_heads_over_limit
)
1999 sc
->gfp_mask
|= __GFP_HIGHMEM
;
2001 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
2002 gfp_zone(sc
->gfp_mask
), sc
->nodemask
) {
2003 if (!populated_zone(zone
))
2006 * Take care memory controller reclaiming has small influence
2009 if (global_reclaim(sc
)) {
2010 if (!cpuset_zone_allowed_hardwall(zone
, GFP_KERNEL
))
2012 if (zone
->all_unreclaimable
&&
2013 sc
->priority
!= DEF_PRIORITY
)
2014 continue; /* Let kswapd poll it */
2015 if (COMPACTION_BUILD
) {
2017 * If we already have plenty of memory free for
2018 * compaction in this zone, don't free any more.
2019 * Even though compaction is invoked for any
2020 * non-zero order, only frequent costly order
2021 * reclamation is disruptive enough to become a
2022 * noticeable problem, like transparent huge
2025 if (compaction_ready(zone
, sc
)) {
2026 aborted_reclaim
= true;
2031 * This steals pages from memory cgroups over softlimit
2032 * and returns the number of reclaimed pages and
2033 * scanned pages. This works for global memory pressure
2034 * and balancing, not for a memcg's limit.
2036 nr_soft_scanned
= 0;
2037 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(zone
,
2038 sc
->order
, sc
->gfp_mask
,
2040 sc
->nr_reclaimed
+= nr_soft_reclaimed
;
2041 sc
->nr_scanned
+= nr_soft_scanned
;
2042 /* need some check for avoid more shrink_zone() */
2045 shrink_zone(zone
, sc
);
2048 return aborted_reclaim
;
2051 static bool zone_reclaimable(struct zone
*zone
)
2053 return zone
->pages_scanned
< zone_reclaimable_pages(zone
) * 6;
2056 /* All zones in zonelist are unreclaimable? */
2057 static bool all_unreclaimable(struct zonelist
*zonelist
,
2058 struct scan_control
*sc
)
2063 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
2064 gfp_zone(sc
->gfp_mask
), sc
->nodemask
) {
2065 if (!populated_zone(zone
))
2067 if (!cpuset_zone_allowed_hardwall(zone
, GFP_KERNEL
))
2069 if (!zone
->all_unreclaimable
)
2077 * This is the main entry point to direct page reclaim.
2079 * If a full scan of the inactive list fails to free enough memory then we
2080 * are "out of memory" and something needs to be killed.
2082 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2083 * high - the zone may be full of dirty or under-writeback pages, which this
2084 * caller can't do much about. We kick the writeback threads and take explicit
2085 * naps in the hope that some of these pages can be written. But if the
2086 * allocating task holds filesystem locks which prevent writeout this might not
2087 * work, and the allocation attempt will fail.
2089 * returns: 0, if no pages reclaimed
2090 * else, the number of pages reclaimed
2092 static unsigned long do_try_to_free_pages(struct zonelist
*zonelist
,
2093 struct scan_control
*sc
,
2094 struct shrink_control
*shrink
)
2096 unsigned long total_scanned
= 0;
2097 struct reclaim_state
*reclaim_state
= current
->reclaim_state
;
2100 unsigned long writeback_threshold
;
2101 bool aborted_reclaim
;
2103 delayacct_freepages_start();
2105 if (global_reclaim(sc
))
2106 count_vm_event(ALLOCSTALL
);
2110 aborted_reclaim
= shrink_zones(zonelist
, sc
);
2113 * Don't shrink slabs when reclaiming memory from
2114 * over limit cgroups
2116 if (global_reclaim(sc
)) {
2117 unsigned long lru_pages
= 0;
2118 for_each_zone_zonelist(zone
, z
, zonelist
,
2119 gfp_zone(sc
->gfp_mask
)) {
2120 if (!cpuset_zone_allowed_hardwall(zone
, GFP_KERNEL
))
2123 lru_pages
+= zone_reclaimable_pages(zone
);
2126 shrink_slab(shrink
, sc
->nr_scanned
, lru_pages
);
2127 if (reclaim_state
) {
2128 sc
->nr_reclaimed
+= reclaim_state
->reclaimed_slab
;
2129 reclaim_state
->reclaimed_slab
= 0;
2132 total_scanned
+= sc
->nr_scanned
;
2133 if (sc
->nr_reclaimed
>= sc
->nr_to_reclaim
)
2137 * Try to write back as many pages as we just scanned. This
2138 * tends to cause slow streaming writers to write data to the
2139 * disk smoothly, at the dirtying rate, which is nice. But
2140 * that's undesirable in laptop mode, where we *want* lumpy
2141 * writeout. So in laptop mode, write out the whole world.
2143 writeback_threshold
= sc
->nr_to_reclaim
+ sc
->nr_to_reclaim
/ 2;
2144 if (total_scanned
> writeback_threshold
) {
2145 wakeup_flusher_threads(laptop_mode
? 0 : total_scanned
,
2146 WB_REASON_TRY_TO_FREE_PAGES
);
2147 sc
->may_writepage
= 1;
2150 /* Take a nap, wait for some writeback to complete */
2151 if (!sc
->hibernation_mode
&& sc
->nr_scanned
&&
2152 sc
->priority
< DEF_PRIORITY
- 2) {
2153 struct zone
*preferred_zone
;
2155 first_zones_zonelist(zonelist
, gfp_zone(sc
->gfp_mask
),
2156 &cpuset_current_mems_allowed
,
2158 wait_iff_congested(preferred_zone
, BLK_RW_ASYNC
, HZ
/10);
2160 } while (--sc
->priority
>= 0);
2163 delayacct_freepages_end();
2165 if (sc
->nr_reclaimed
)
2166 return sc
->nr_reclaimed
;
2169 * As hibernation is going on, kswapd is freezed so that it can't mark
2170 * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2173 if (oom_killer_disabled
)
2176 /* Aborted reclaim to try compaction? don't OOM, then */
2177 if (aborted_reclaim
)
2180 /* top priority shrink_zones still had more to do? don't OOM, then */
2181 if (global_reclaim(sc
) && !all_unreclaimable(zonelist
, sc
))
2187 unsigned long try_to_free_pages(struct zonelist
*zonelist
, int order
,
2188 gfp_t gfp_mask
, nodemask_t
*nodemask
)
2190 unsigned long nr_reclaimed
;
2191 struct scan_control sc
= {
2192 .gfp_mask
= gfp_mask
,
2193 .may_writepage
= !laptop_mode
,
2194 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
2198 .priority
= DEF_PRIORITY
,
2199 .target_mem_cgroup
= NULL
,
2200 .nodemask
= nodemask
,
2202 struct shrink_control shrink
= {
2203 .gfp_mask
= sc
.gfp_mask
,
2206 trace_mm_vmscan_direct_reclaim_begin(order
,
2210 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
, &shrink
);
2212 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed
);
2214 return nr_reclaimed
;
2217 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
2219 unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup
*memcg
,
2220 gfp_t gfp_mask
, bool noswap
,
2222 unsigned long *nr_scanned
)
2224 struct scan_control sc
= {
2226 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
2227 .may_writepage
= !laptop_mode
,
2229 .may_swap
= !noswap
,
2232 .target_mem_cgroup
= memcg
,
2234 struct mem_cgroup_zone mz
= {
2235 .mem_cgroup
= memcg
,
2239 sc
.gfp_mask
= (gfp_mask
& GFP_RECLAIM_MASK
) |
2240 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
);
2242 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc
.order
,
2247 * NOTE: Although we can get the priority field, using it
2248 * here is not a good idea, since it limits the pages we can scan.
2249 * if we don't reclaim here, the shrink_zone from balance_pgdat
2250 * will pick up pages from other mem cgroup's as well. We hack
2251 * the priority and make it zero.
2253 shrink_mem_cgroup_zone(&mz
, &sc
);
2255 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc
.nr_reclaimed
);
2257 *nr_scanned
= sc
.nr_scanned
;
2258 return sc
.nr_reclaimed
;
2261 unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup
*memcg
,
2265 struct zonelist
*zonelist
;
2266 unsigned long nr_reclaimed
;
2268 struct scan_control sc
= {
2269 .may_writepage
= !laptop_mode
,
2271 .may_swap
= !noswap
,
2272 .nr_to_reclaim
= SWAP_CLUSTER_MAX
,
2274 .priority
= DEF_PRIORITY
,
2275 .target_mem_cgroup
= memcg
,
2276 .nodemask
= NULL
, /* we don't care the placement */
2277 .gfp_mask
= (gfp_mask
& GFP_RECLAIM_MASK
) |
2278 (GFP_HIGHUSER_MOVABLE
& ~GFP_RECLAIM_MASK
),
2280 struct shrink_control shrink
= {
2281 .gfp_mask
= sc
.gfp_mask
,
2285 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2286 * take care of from where we get pages. So the node where we start the
2287 * scan does not need to be the current node.
2289 nid
= mem_cgroup_select_victim_node(memcg
);
2291 zonelist
= NODE_DATA(nid
)->node_zonelists
;
2293 trace_mm_vmscan_memcg_reclaim_begin(0,
2297 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
, &shrink
);
2299 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed
);
2301 return nr_reclaimed
;
2305 static void age_active_anon(struct zone
*zone
, struct scan_control
*sc
)
2307 struct mem_cgroup
*memcg
;
2309 if (!total_swap_pages
)
2312 memcg
= mem_cgroup_iter(NULL
, NULL
, NULL
);
2314 struct mem_cgroup_zone mz
= {
2315 .mem_cgroup
= memcg
,
2319 if (inactive_anon_is_low(&mz
))
2320 shrink_active_list(SWAP_CLUSTER_MAX
, &mz
,
2321 sc
, LRU_ACTIVE_ANON
);
2323 memcg
= mem_cgroup_iter(NULL
, memcg
, NULL
);
2328 * pgdat_balanced is used when checking if a node is balanced for high-order
2329 * allocations. Only zones that meet watermarks and are in a zone allowed
2330 * by the callers classzone_idx are added to balanced_pages. The total of
2331 * balanced pages must be at least 25% of the zones allowed by classzone_idx
2332 * for the node to be considered balanced. Forcing all zones to be balanced
2333 * for high orders can cause excessive reclaim when there are imbalanced zones.
2334 * The choice of 25% is due to
2335 * o a 16M DMA zone that is balanced will not balance a zone on any
2336 * reasonable sized machine
2337 * o On all other machines, the top zone must be at least a reasonable
2338 * percentage of the middle zones. For example, on 32-bit x86, highmem
2339 * would need to be at least 256M for it to be balance a whole node.
2340 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2341 * to balance a node on its own. These seemed like reasonable ratios.
2343 static bool pgdat_balanced(pg_data_t
*pgdat
, unsigned long balanced_pages
,
2346 unsigned long present_pages
= 0;
2349 for (i
= 0; i
<= classzone_idx
; i
++)
2350 present_pages
+= pgdat
->node_zones
[i
].present_pages
;
2352 /* A special case here: if zone has no page, we think it's balanced */
2353 return balanced_pages
>= (present_pages
>> 2);
2356 /* is kswapd sleeping prematurely? */
2357 static bool sleeping_prematurely(pg_data_t
*pgdat
, int order
, long remaining
,
2361 unsigned long balanced
= 0;
2362 bool all_zones_ok
= true;
2364 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2368 /* Check the watermark levels */
2369 for (i
= 0; i
<= classzone_idx
; i
++) {
2370 struct zone
*zone
= pgdat
->node_zones
+ i
;
2372 if (!populated_zone(zone
))
2376 * balance_pgdat() skips over all_unreclaimable after
2377 * DEF_PRIORITY. Effectively, it considers them balanced so
2378 * they must be considered balanced here as well if kswapd
2381 if (zone
->all_unreclaimable
) {
2382 balanced
+= zone
->present_pages
;
2386 if (!zone_watermark_ok_safe(zone
, order
, high_wmark_pages(zone
),
2388 all_zones_ok
= false;
2390 balanced
+= zone
->present_pages
;
2394 * For high-order requests, the balanced zones must contain at least
2395 * 25% of the nodes pages for kswapd to sleep. For order-0, all zones
2399 return !pgdat_balanced(pgdat
, balanced
, classzone_idx
);
2401 return !all_zones_ok
;
2405 * For kswapd, balance_pgdat() will work across all this node's zones until
2406 * they are all at high_wmark_pages(zone).
2408 * Returns the final order kswapd was reclaiming at
2410 * There is special handling here for zones which are full of pinned pages.
2411 * This can happen if the pages are all mlocked, or if they are all used by
2412 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
2413 * What we do is to detect the case where all pages in the zone have been
2414 * scanned twice and there has been zero successful reclaim. Mark the zone as
2415 * dead and from now on, only perform a short scan. Basically we're polling
2416 * the zone for when the problem goes away.
2418 * kswapd scans the zones in the highmem->normal->dma direction. It skips
2419 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2420 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2421 * lower zones regardless of the number of free pages in the lower zones. This
2422 * interoperates with the page allocator fallback scheme to ensure that aging
2423 * of pages is balanced across the zones.
2425 static unsigned long balance_pgdat(pg_data_t
*pgdat
, int order
,
2429 unsigned long balanced
;
2431 int end_zone
= 0; /* Inclusive. 0 = ZONE_DMA */
2432 unsigned long total_scanned
;
2433 struct reclaim_state
*reclaim_state
= current
->reclaim_state
;
2434 unsigned long nr_soft_reclaimed
;
2435 unsigned long nr_soft_scanned
;
2436 struct scan_control sc
= {
2437 .gfp_mask
= GFP_KERNEL
,
2441 * kswapd doesn't want to be bailed out while reclaim. because
2442 * we want to put equal scanning pressure on each zone.
2444 .nr_to_reclaim
= ULONG_MAX
,
2446 .target_mem_cgroup
= NULL
,
2448 struct shrink_control shrink
= {
2449 .gfp_mask
= sc
.gfp_mask
,
2453 sc
.priority
= DEF_PRIORITY
;
2454 sc
.nr_reclaimed
= 0;
2455 sc
.may_writepage
= !laptop_mode
;
2456 count_vm_event(PAGEOUTRUN
);
2459 unsigned long lru_pages
= 0;
2460 int has_under_min_watermark_zone
= 0;
2466 * Scan in the highmem->dma direction for the highest
2467 * zone which needs scanning
2469 for (i
= pgdat
->nr_zones
- 1; i
>= 0; i
--) {
2470 struct zone
*zone
= pgdat
->node_zones
+ i
;
2472 if (!populated_zone(zone
))
2475 if (zone
->all_unreclaimable
&&
2476 sc
.priority
!= DEF_PRIORITY
)
2480 * Do some background aging of the anon list, to give
2481 * pages a chance to be referenced before reclaiming.
2483 age_active_anon(zone
, &sc
);
2486 * If the number of buffer_heads in the machine
2487 * exceeds the maximum allowed level and this node
2488 * has a highmem zone, force kswapd to reclaim from
2489 * it to relieve lowmem pressure.
2491 if (buffer_heads_over_limit
&& is_highmem_idx(i
)) {
2496 if (!zone_watermark_ok_safe(zone
, order
,
2497 high_wmark_pages(zone
), 0, 0)) {
2501 /* If balanced, clear the congested flag */
2502 zone_clear_flag(zone
, ZONE_CONGESTED
);
2508 for (i
= 0; i
<= end_zone
; i
++) {
2509 struct zone
*zone
= pgdat
->node_zones
+ i
;
2511 lru_pages
+= zone_reclaimable_pages(zone
);
2515 * Now scan the zone in the dma->highmem direction, stopping
2516 * at the last zone which needs scanning.
2518 * We do this because the page allocator works in the opposite
2519 * direction. This prevents the page allocator from allocating
2520 * pages behind kswapd's direction of progress, which would
2521 * cause too much scanning of the lower zones.
2523 for (i
= 0; i
<= end_zone
; i
++) {
2524 struct zone
*zone
= pgdat
->node_zones
+ i
;
2525 int nr_slab
, testorder
;
2526 unsigned long balance_gap
;
2528 if (!populated_zone(zone
))
2531 if (zone
->all_unreclaimable
&&
2532 sc
.priority
!= DEF_PRIORITY
)
2537 nr_soft_scanned
= 0;
2539 * Call soft limit reclaim before calling shrink_zone.
2541 nr_soft_reclaimed
= mem_cgroup_soft_limit_reclaim(zone
,
2544 sc
.nr_reclaimed
+= nr_soft_reclaimed
;
2545 total_scanned
+= nr_soft_scanned
;
2548 * We put equal pressure on every zone, unless
2549 * one zone has way too many pages free
2550 * already. The "too many pages" is defined
2551 * as the high wmark plus a "gap" where the
2552 * gap is either the low watermark or 1%
2553 * of the zone, whichever is smaller.
2555 balance_gap
= min(low_wmark_pages(zone
),
2556 (zone
->present_pages
+
2557 KSWAPD_ZONE_BALANCE_GAP_RATIO
-1) /
2558 KSWAPD_ZONE_BALANCE_GAP_RATIO
);
2560 * Kswapd reclaims only single pages with compaction
2561 * enabled. Trying too hard to reclaim until contiguous
2562 * free pages have become available can hurt performance
2563 * by evicting too much useful data from memory.
2564 * Do not reclaim more than needed for compaction.
2567 if (COMPACTION_BUILD
&& order
&&
2568 compaction_suitable(zone
, order
) !=
2572 if ((buffer_heads_over_limit
&& is_highmem_idx(i
)) ||
2573 !zone_watermark_ok_safe(zone
, testorder
,
2574 high_wmark_pages(zone
) + balance_gap
,
2576 shrink_zone(zone
, &sc
);
2578 reclaim_state
->reclaimed_slab
= 0;
2579 nr_slab
= shrink_slab(&shrink
, sc
.nr_scanned
, lru_pages
);
2580 sc
.nr_reclaimed
+= reclaim_state
->reclaimed_slab
;
2581 total_scanned
+= sc
.nr_scanned
;
2583 if (nr_slab
== 0 && !zone_reclaimable(zone
))
2584 zone
->all_unreclaimable
= 1;
2588 * If we've done a decent amount of scanning and
2589 * the reclaim ratio is low, start doing writepage
2590 * even in laptop mode
2592 if (total_scanned
> SWAP_CLUSTER_MAX
* 2 &&
2593 total_scanned
> sc
.nr_reclaimed
+ sc
.nr_reclaimed
/ 2)
2594 sc
.may_writepage
= 1;
2596 if (zone
->all_unreclaimable
) {
2597 if (end_zone
&& end_zone
== i
)
2602 if (!zone_watermark_ok_safe(zone
, testorder
,
2603 high_wmark_pages(zone
), end_zone
, 0)) {
2606 * We are still under min water mark. This
2607 * means that we have a GFP_ATOMIC allocation
2608 * failure risk. Hurry up!
2610 if (!zone_watermark_ok_safe(zone
, order
,
2611 min_wmark_pages(zone
), end_zone
, 0))
2612 has_under_min_watermark_zone
= 1;
2615 * If a zone reaches its high watermark,
2616 * consider it to be no longer congested. It's
2617 * possible there are dirty pages backed by
2618 * congested BDIs but as pressure is relieved,
2619 * spectulatively avoid congestion waits
2621 zone_clear_flag(zone
, ZONE_CONGESTED
);
2622 if (i
<= *classzone_idx
)
2623 balanced
+= zone
->present_pages
;
2627 if (all_zones_ok
|| (order
&& pgdat_balanced(pgdat
, balanced
, *classzone_idx
)))
2628 break; /* kswapd: all done */
2630 * OK, kswapd is getting into trouble. Take a nap, then take
2631 * another pass across the zones.
2633 if (total_scanned
&& (sc
.priority
< DEF_PRIORITY
- 2)) {
2634 if (has_under_min_watermark_zone
)
2635 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT
);
2637 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
2641 * We do this so kswapd doesn't build up large priorities for
2642 * example when it is freeing in parallel with allocators. It
2643 * matches the direct reclaim path behaviour in terms of impact
2644 * on zone->*_priority.
2646 if (sc
.nr_reclaimed
>= SWAP_CLUSTER_MAX
)
2648 } while (--sc
.priority
>= 0);
2652 * order-0: All zones must meet high watermark for a balanced node
2653 * high-order: Balanced zones must make up at least 25% of the node
2654 * for the node to be balanced
2656 if (!(all_zones_ok
|| (order
&& pgdat_balanced(pgdat
, balanced
, *classzone_idx
)))) {
2662 * Fragmentation may mean that the system cannot be
2663 * rebalanced for high-order allocations in all zones.
2664 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2665 * it means the zones have been fully scanned and are still
2666 * not balanced. For high-order allocations, there is
2667 * little point trying all over again as kswapd may
2670 * Instead, recheck all watermarks at order-0 as they
2671 * are the most important. If watermarks are ok, kswapd will go
2672 * back to sleep. High-order users can still perform direct
2673 * reclaim if they wish.
2675 if (sc
.nr_reclaimed
< SWAP_CLUSTER_MAX
)
2676 order
= sc
.order
= 0;
2682 * If kswapd was reclaiming at a higher order, it has the option of
2683 * sleeping without all zones being balanced. Before it does, it must
2684 * ensure that the watermarks for order-0 on *all* zones are met and
2685 * that the congestion flags are cleared. The congestion flag must
2686 * be cleared as kswapd is the only mechanism that clears the flag
2687 * and it is potentially going to sleep here.
2690 int zones_need_compaction
= 1;
2692 for (i
= 0; i
<= end_zone
; i
++) {
2693 struct zone
*zone
= pgdat
->node_zones
+ i
;
2695 if (!populated_zone(zone
))
2698 if (zone
->all_unreclaimable
&&
2699 sc
.priority
!= DEF_PRIORITY
)
2702 /* Would compaction fail due to lack of free memory? */
2703 if (COMPACTION_BUILD
&&
2704 compaction_suitable(zone
, order
) == COMPACT_SKIPPED
)
2707 /* Confirm the zone is balanced for order-0 */
2708 if (!zone_watermark_ok(zone
, 0,
2709 high_wmark_pages(zone
), 0, 0)) {
2710 order
= sc
.order
= 0;
2714 /* Check if the memory needs to be defragmented. */
2715 if (zone_watermark_ok(zone
, order
,
2716 low_wmark_pages(zone
), *classzone_idx
, 0))
2717 zones_need_compaction
= 0;
2719 /* If balanced, clear the congested flag */
2720 zone_clear_flag(zone
, ZONE_CONGESTED
);
2723 if (zones_need_compaction
)
2724 compact_pgdat(pgdat
, order
);
2728 * Return the order we were reclaiming at so sleeping_prematurely()
2729 * makes a decision on the order we were last reclaiming at. However,
2730 * if another caller entered the allocator slow path while kswapd
2731 * was awake, order will remain at the higher level
2733 *classzone_idx
= end_zone
;
2737 static void kswapd_try_to_sleep(pg_data_t
*pgdat
, int order
, int classzone_idx
)
2742 if (freezing(current
) || kthread_should_stop())
2745 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
2747 /* Try to sleep for a short interval */
2748 if (!sleeping_prematurely(pgdat
, order
, remaining
, classzone_idx
)) {
2749 remaining
= schedule_timeout(HZ
/10);
2750 finish_wait(&pgdat
->kswapd_wait
, &wait
);
2751 prepare_to_wait(&pgdat
->kswapd_wait
, &wait
, TASK_INTERRUPTIBLE
);
2755 * After a short sleep, check if it was a premature sleep. If not, then
2756 * go fully to sleep until explicitly woken up.
2758 if (!sleeping_prematurely(pgdat
, order
, remaining
, classzone_idx
)) {
2759 trace_mm_vmscan_kswapd_sleep(pgdat
->node_id
);
2762 * vmstat counters are not perfectly accurate and the estimated
2763 * value for counters such as NR_FREE_PAGES can deviate from the
2764 * true value by nr_online_cpus * threshold. To avoid the zone
2765 * watermarks being breached while under pressure, we reduce the
2766 * per-cpu vmstat threshold while kswapd is awake and restore
2767 * them before going back to sleep.
2769 set_pgdat_percpu_threshold(pgdat
, calculate_normal_threshold
);
2771 set_pgdat_percpu_threshold(pgdat
, calculate_pressure_threshold
);
2774 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY
);
2776 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY
);
2778 finish_wait(&pgdat
->kswapd_wait
, &wait
);
2782 * The background pageout daemon, started as a kernel thread
2783 * from the init process.
2785 * This basically trickles out pages so that we have _some_
2786 * free memory available even if there is no other activity
2787 * that frees anything up. This is needed for things like routing
2788 * etc, where we otherwise might have all activity going on in
2789 * asynchronous contexts that cannot page things out.
2791 * If there are applications that are active memory-allocators
2792 * (most normal use), this basically shouldn't matter.
2794 static int kswapd(void *p
)
2796 unsigned long order
, new_order
;
2797 unsigned balanced_order
;
2798 int classzone_idx
, new_classzone_idx
;
2799 int balanced_classzone_idx
;
2800 pg_data_t
*pgdat
= (pg_data_t
*)p
;
2801 struct task_struct
*tsk
= current
;
2803 struct reclaim_state reclaim_state
= {
2804 .reclaimed_slab
= 0,
2806 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
2808 lockdep_set_current_reclaim_state(GFP_KERNEL
);
2810 if (!cpumask_empty(cpumask
))
2811 set_cpus_allowed_ptr(tsk
, cpumask
);
2812 current
->reclaim_state
= &reclaim_state
;
2815 * Tell the memory management that we're a "memory allocator",
2816 * and that if we need more memory we should get access to it
2817 * regardless (see "__alloc_pages()"). "kswapd" should
2818 * never get caught in the normal page freeing logic.
2820 * (Kswapd normally doesn't need memory anyway, but sometimes
2821 * you need a small amount of memory in order to be able to
2822 * page out something else, and this flag essentially protects
2823 * us from recursively trying to free more memory as we're
2824 * trying to free the first piece of memory in the first place).
2826 tsk
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
| PF_KSWAPD
;
2829 order
= new_order
= 0;
2831 classzone_idx
= new_classzone_idx
= pgdat
->nr_zones
- 1;
2832 balanced_classzone_idx
= classzone_idx
;
2837 * If the last balance_pgdat was unsuccessful it's unlikely a
2838 * new request of a similar or harder type will succeed soon
2839 * so consider going to sleep on the basis we reclaimed at
2841 if (balanced_classzone_idx
>= new_classzone_idx
&&
2842 balanced_order
== new_order
) {
2843 new_order
= pgdat
->kswapd_max_order
;
2844 new_classzone_idx
= pgdat
->classzone_idx
;
2845 pgdat
->kswapd_max_order
= 0;
2846 pgdat
->classzone_idx
= pgdat
->nr_zones
- 1;
2849 if (order
< new_order
|| classzone_idx
> new_classzone_idx
) {
2851 * Don't sleep if someone wants a larger 'order'
2852 * allocation or has tigher zone constraints
2855 classzone_idx
= new_classzone_idx
;
2857 kswapd_try_to_sleep(pgdat
, balanced_order
,
2858 balanced_classzone_idx
);
2859 order
= pgdat
->kswapd_max_order
;
2860 classzone_idx
= pgdat
->classzone_idx
;
2862 new_classzone_idx
= classzone_idx
;
2863 pgdat
->kswapd_max_order
= 0;
2864 pgdat
->classzone_idx
= pgdat
->nr_zones
- 1;
2867 ret
= try_to_freeze();
2868 if (kthread_should_stop())
2872 * We can speed up thawing tasks if we don't call balance_pgdat
2873 * after returning from the refrigerator
2876 trace_mm_vmscan_kswapd_wake(pgdat
->node_id
, order
);
2877 balanced_classzone_idx
= classzone_idx
;
2878 balanced_order
= balance_pgdat(pgdat
, order
,
2879 &balanced_classzone_idx
);
2886 * A zone is low on free memory, so wake its kswapd task to service it.
2888 void wakeup_kswapd(struct zone
*zone
, int order
, enum zone_type classzone_idx
)
2892 if (!populated_zone(zone
))
2895 if (!cpuset_zone_allowed_hardwall(zone
, GFP_KERNEL
))
2897 pgdat
= zone
->zone_pgdat
;
2898 if (pgdat
->kswapd_max_order
< order
) {
2899 pgdat
->kswapd_max_order
= order
;
2900 pgdat
->classzone_idx
= min(pgdat
->classzone_idx
, classzone_idx
);
2902 if (!waitqueue_active(&pgdat
->kswapd_wait
))
2904 if (zone_watermark_ok_safe(zone
, order
, low_wmark_pages(zone
), 0, 0))
2907 trace_mm_vmscan_wakeup_kswapd(pgdat
->node_id
, zone_idx(zone
), order
);
2908 wake_up_interruptible(&pgdat
->kswapd_wait
);
2912 * The reclaimable count would be mostly accurate.
2913 * The less reclaimable pages may be
2914 * - mlocked pages, which will be moved to unevictable list when encountered
2915 * - mapped pages, which may require several travels to be reclaimed
2916 * - dirty pages, which is not "instantly" reclaimable
2918 unsigned long global_reclaimable_pages(void)
2922 nr
= global_page_state(NR_ACTIVE_FILE
) +
2923 global_page_state(NR_INACTIVE_FILE
);
2925 if (nr_swap_pages
> 0)
2926 nr
+= global_page_state(NR_ACTIVE_ANON
) +
2927 global_page_state(NR_INACTIVE_ANON
);
2932 unsigned long zone_reclaimable_pages(struct zone
*zone
)
2936 nr
= zone_page_state(zone
, NR_ACTIVE_FILE
) +
2937 zone_page_state(zone
, NR_INACTIVE_FILE
);
2939 if (nr_swap_pages
> 0)
2940 nr
+= zone_page_state(zone
, NR_ACTIVE_ANON
) +
2941 zone_page_state(zone
, NR_INACTIVE_ANON
);
2946 #ifdef CONFIG_HIBERNATION
2948 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
2951 * Rather than trying to age LRUs the aim is to preserve the overall
2952 * LRU order by reclaiming preferentially
2953 * inactive > active > active referenced > active mapped
2955 unsigned long shrink_all_memory(unsigned long nr_to_reclaim
)
2957 struct reclaim_state reclaim_state
;
2958 struct scan_control sc
= {
2959 .gfp_mask
= GFP_HIGHUSER_MOVABLE
,
2963 .nr_to_reclaim
= nr_to_reclaim
,
2964 .hibernation_mode
= 1,
2966 .priority
= DEF_PRIORITY
,
2968 struct shrink_control shrink
= {
2969 .gfp_mask
= sc
.gfp_mask
,
2971 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), sc
.gfp_mask
);
2972 struct task_struct
*p
= current
;
2973 unsigned long nr_reclaimed
;
2975 p
->flags
|= PF_MEMALLOC
;
2976 lockdep_set_current_reclaim_state(sc
.gfp_mask
);
2977 reclaim_state
.reclaimed_slab
= 0;
2978 p
->reclaim_state
= &reclaim_state
;
2980 nr_reclaimed
= do_try_to_free_pages(zonelist
, &sc
, &shrink
);
2982 p
->reclaim_state
= NULL
;
2983 lockdep_clear_current_reclaim_state();
2984 p
->flags
&= ~PF_MEMALLOC
;
2986 return nr_reclaimed
;
2988 #endif /* CONFIG_HIBERNATION */
2990 /* It's optimal to keep kswapds on the same CPUs as their memory, but
2991 not required for correctness. So if the last cpu in a node goes
2992 away, we get changed to run anywhere: as the first one comes back,
2993 restore their cpu bindings. */
2994 static int __devinit
cpu_callback(struct notifier_block
*nfb
,
2995 unsigned long action
, void *hcpu
)
2999 if (action
== CPU_ONLINE
|| action
== CPU_ONLINE_FROZEN
) {
3000 for_each_node_state(nid
, N_HIGH_MEMORY
) {
3001 pg_data_t
*pgdat
= NODE_DATA(nid
);
3002 const struct cpumask
*mask
;
3004 mask
= cpumask_of_node(pgdat
->node_id
);
3006 if (cpumask_any_and(cpu_online_mask
, mask
) < nr_cpu_ids
)
3007 /* One of our CPUs online: restore mask */
3008 set_cpus_allowed_ptr(pgdat
->kswapd
, mask
);
3015 * This kswapd start function will be called by init and node-hot-add.
3016 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3018 int kswapd_run(int nid
)
3020 pg_data_t
*pgdat
= NODE_DATA(nid
);
3026 pgdat
->kswapd
= kthread_run(kswapd
, pgdat
, "kswapd%d", nid
);
3027 if (IS_ERR(pgdat
->kswapd
)) {
3028 /* failure at boot is fatal */
3029 BUG_ON(system_state
== SYSTEM_BOOTING
);
3030 printk("Failed to start kswapd on node %d\n",nid
);
3037 * Called by memory hotplug when all memory in a node is offlined.
3039 void kswapd_stop(int nid
)
3041 struct task_struct
*kswapd
= NODE_DATA(nid
)->kswapd
;
3044 kthread_stop(kswapd
);
3047 static int __init
kswapd_init(void)
3052 for_each_node_state(nid
, N_HIGH_MEMORY
)
3054 hotcpu_notifier(cpu_callback
, 0);
3058 module_init(kswapd_init
)
3064 * If non-zero call zone_reclaim when the number of free pages falls below
3067 int zone_reclaim_mode __read_mostly
;
3069 #define RECLAIM_OFF 0
3070 #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
3071 #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3072 #define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
3075 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3076 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3079 #define ZONE_RECLAIM_PRIORITY 4
3082 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3085 int sysctl_min_unmapped_ratio
= 1;
3088 * If the number of slab pages in a zone grows beyond this percentage then
3089 * slab reclaim needs to occur.
3091 int sysctl_min_slab_ratio
= 5;
3093 static inline unsigned long zone_unmapped_file_pages(struct zone
*zone
)
3095 unsigned long file_mapped
= zone_page_state(zone
, NR_FILE_MAPPED
);
3096 unsigned long file_lru
= zone_page_state(zone
, NR_INACTIVE_FILE
) +
3097 zone_page_state(zone
, NR_ACTIVE_FILE
);
3100 * It's possible for there to be more file mapped pages than
3101 * accounted for by the pages on the file LRU lists because
3102 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3104 return (file_lru
> file_mapped
) ? (file_lru
- file_mapped
) : 0;
3107 /* Work out how many page cache pages we can reclaim in this reclaim_mode */
3108 static long zone_pagecache_reclaimable(struct zone
*zone
)
3110 long nr_pagecache_reclaimable
;
3114 * If RECLAIM_SWAP is set, then all file pages are considered
3115 * potentially reclaimable. Otherwise, we have to worry about
3116 * pages like swapcache and zone_unmapped_file_pages() provides
3119 if (zone_reclaim_mode
& RECLAIM_SWAP
)
3120 nr_pagecache_reclaimable
= zone_page_state(zone
, NR_FILE_PAGES
);
3122 nr_pagecache_reclaimable
= zone_unmapped_file_pages(zone
);
3124 /* If we can't clean pages, remove dirty pages from consideration */
3125 if (!(zone_reclaim_mode
& RECLAIM_WRITE
))
3126 delta
+= zone_page_state(zone
, NR_FILE_DIRTY
);
3128 /* Watch for any possible underflows due to delta */
3129 if (unlikely(delta
> nr_pagecache_reclaimable
))
3130 delta
= nr_pagecache_reclaimable
;
3132 return nr_pagecache_reclaimable
- delta
;
3136 * Try to free up some pages from this zone through reclaim.
3138 static int __zone_reclaim(struct zone
*zone
, gfp_t gfp_mask
, unsigned int order
)
3140 /* Minimum pages needed in order to stay on node */
3141 const unsigned long nr_pages
= 1 << order
;
3142 struct task_struct
*p
= current
;
3143 struct reclaim_state reclaim_state
;
3144 struct scan_control sc
= {
3145 .may_writepage
= !!(zone_reclaim_mode
& RECLAIM_WRITE
),
3146 .may_unmap
= !!(zone_reclaim_mode
& RECLAIM_SWAP
),
3148 .nr_to_reclaim
= max_t(unsigned long, nr_pages
,
3150 .gfp_mask
= gfp_mask
,
3152 .priority
= ZONE_RECLAIM_PRIORITY
,
3154 struct shrink_control shrink
= {
3155 .gfp_mask
= sc
.gfp_mask
,
3157 unsigned long nr_slab_pages0
, nr_slab_pages1
;
3161 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3162 * and we also need to be able to write out pages for RECLAIM_WRITE
3165 p
->flags
|= PF_MEMALLOC
| PF_SWAPWRITE
;
3166 lockdep_set_current_reclaim_state(gfp_mask
);
3167 reclaim_state
.reclaimed_slab
= 0;
3168 p
->reclaim_state
= &reclaim_state
;
3170 if (zone_pagecache_reclaimable(zone
) > zone
->min_unmapped_pages
) {
3172 * Free memory by calling shrink zone with increasing
3173 * priorities until we have enough memory freed.
3176 shrink_zone(zone
, &sc
);
3177 } while (sc
.nr_reclaimed
< nr_pages
&& --sc
.priority
>= 0);
3180 nr_slab_pages0
= zone_page_state(zone
, NR_SLAB_RECLAIMABLE
);
3181 if (nr_slab_pages0
> zone
->min_slab_pages
) {
3183 * shrink_slab() does not currently allow us to determine how
3184 * many pages were freed in this zone. So we take the current
3185 * number of slab pages and shake the slab until it is reduced
3186 * by the same nr_pages that we used for reclaiming unmapped
3189 * Note that shrink_slab will free memory on all zones and may
3193 unsigned long lru_pages
= zone_reclaimable_pages(zone
);
3195 /* No reclaimable slab or very low memory pressure */
3196 if (!shrink_slab(&shrink
, sc
.nr_scanned
, lru_pages
))
3199 /* Freed enough memory */
3200 nr_slab_pages1
= zone_page_state(zone
,
3201 NR_SLAB_RECLAIMABLE
);
3202 if (nr_slab_pages1
+ nr_pages
<= nr_slab_pages0
)
3207 * Update nr_reclaimed by the number of slab pages we
3208 * reclaimed from this zone.
3210 nr_slab_pages1
= zone_page_state(zone
, NR_SLAB_RECLAIMABLE
);
3211 if (nr_slab_pages1
< nr_slab_pages0
)
3212 sc
.nr_reclaimed
+= nr_slab_pages0
- nr_slab_pages1
;
3215 p
->reclaim_state
= NULL
;
3216 current
->flags
&= ~(PF_MEMALLOC
| PF_SWAPWRITE
);
3217 lockdep_clear_current_reclaim_state();
3218 return sc
.nr_reclaimed
>= nr_pages
;
3221 int zone_reclaim(struct zone
*zone
, gfp_t gfp_mask
, unsigned int order
)
3227 * Zone reclaim reclaims unmapped file backed pages and
3228 * slab pages if we are over the defined limits.
3230 * A small portion of unmapped file backed pages is needed for
3231 * file I/O otherwise pages read by file I/O will be immediately
3232 * thrown out if the zone is overallocated. So we do not reclaim
3233 * if less than a specified percentage of the zone is used by
3234 * unmapped file backed pages.
3236 if (zone_pagecache_reclaimable(zone
) <= zone
->min_unmapped_pages
&&
3237 zone_page_state(zone
, NR_SLAB_RECLAIMABLE
) <= zone
->min_slab_pages
)
3238 return ZONE_RECLAIM_FULL
;
3240 if (zone
->all_unreclaimable
)
3241 return ZONE_RECLAIM_FULL
;
3244 * Do not scan if the allocation should not be delayed.
3246 if (!(gfp_mask
& __GFP_WAIT
) || (current
->flags
& PF_MEMALLOC
))
3247 return ZONE_RECLAIM_NOSCAN
;
3250 * Only run zone reclaim on the local zone or on zones that do not
3251 * have associated processors. This will favor the local processor
3252 * over remote processors and spread off node memory allocations
3253 * as wide as possible.
3255 node_id
= zone_to_nid(zone
);
3256 if (node_state(node_id
, N_CPU
) && node_id
!= numa_node_id())
3257 return ZONE_RECLAIM_NOSCAN
;
3259 if (zone_test_and_set_flag(zone
, ZONE_RECLAIM_LOCKED
))
3260 return ZONE_RECLAIM_NOSCAN
;
3262 ret
= __zone_reclaim(zone
, gfp_mask
, order
);
3263 zone_clear_flag(zone
, ZONE_RECLAIM_LOCKED
);
3266 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED
);
3273 * page_evictable - test whether a page is evictable
3274 * @page: the page to test
3275 * @vma: the VMA in which the page is or will be mapped, may be NULL
3277 * Test whether page is evictable--i.e., should be placed on active/inactive
3278 * lists vs unevictable list. The vma argument is !NULL when called from the
3279 * fault path to determine how to instantate a new page.
3281 * Reasons page might not be evictable:
3282 * (1) page's mapping marked unevictable
3283 * (2) page is part of an mlocked VMA
3286 int page_evictable(struct page
*page
, struct vm_area_struct
*vma
)
3289 if (mapping_unevictable(page_mapping(page
)))
3292 if (PageMlocked(page
) || (vma
&& mlocked_vma_newpage(vma
, page
)))
3300 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3301 * @pages: array of pages to check
3302 * @nr_pages: number of pages to check
3304 * Checks pages for evictability and moves them to the appropriate lru list.
3306 * This function is only used for SysV IPC SHM_UNLOCK.
3308 void check_move_unevictable_pages(struct page
**pages
, int nr_pages
)
3310 struct lruvec
*lruvec
;
3311 struct zone
*zone
= NULL
;
3316 for (i
= 0; i
< nr_pages
; i
++) {
3317 struct page
*page
= pages
[i
];
3318 struct zone
*pagezone
;
3321 pagezone
= page_zone(page
);
3322 if (pagezone
!= zone
) {
3324 spin_unlock_irq(&zone
->lru_lock
);
3326 spin_lock_irq(&zone
->lru_lock
);
3329 if (!PageLRU(page
) || !PageUnevictable(page
))
3332 if (page_evictable(page
, NULL
)) {
3333 enum lru_list lru
= page_lru_base_type(page
);
3335 VM_BUG_ON(PageActive(page
));
3336 ClearPageUnevictable(page
);
3337 __dec_zone_state(zone
, NR_UNEVICTABLE
);
3338 lruvec
= mem_cgroup_lru_move_lists(zone
, page
,
3339 LRU_UNEVICTABLE
, lru
);
3340 list_move(&page
->lru
, &lruvec
->lists
[lru
]);
3341 __inc_zone_state(zone
, NR_INACTIVE_ANON
+ lru
);
3347 __count_vm_events(UNEVICTABLE_PGRESCUED
, pgrescued
);
3348 __count_vm_events(UNEVICTABLE_PGSCANNED
, pgscanned
);
3349 spin_unlock_irq(&zone
->lru_lock
);
3352 #endif /* CONFIG_SHMEM */
3354 static void warn_scan_unevictable_pages(void)
3356 printk_once(KERN_WARNING
3357 "%s: The scan_unevictable_pages sysctl/node-interface has been "
3358 "disabled for lack of a legitimate use case. If you have "
3359 "one, please send an email to linux-mm@kvack.org.\n",
3364 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3365 * all nodes' unevictable lists for evictable pages
3367 unsigned long scan_unevictable_pages
;
3369 int scan_unevictable_handler(struct ctl_table
*table
, int write
,
3370 void __user
*buffer
,
3371 size_t *length
, loff_t
*ppos
)
3373 warn_scan_unevictable_pages();
3374 proc_doulongvec_minmax(table
, write
, buffer
, length
, ppos
);
3375 scan_unevictable_pages
= 0;
3381 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3382 * a specified node's per zone unevictable lists for evictable pages.
3385 static ssize_t
read_scan_unevictable_node(struct device
*dev
,
3386 struct device_attribute
*attr
,
3389 warn_scan_unevictable_pages();
3390 return sprintf(buf
, "0\n"); /* always zero; should fit... */
3393 static ssize_t
write_scan_unevictable_node(struct device
*dev
,
3394 struct device_attribute
*attr
,
3395 const char *buf
, size_t count
)
3397 warn_scan_unevictable_pages();
3402 static DEVICE_ATTR(scan_unevictable_pages
, S_IRUGO
| S_IWUSR
,
3403 read_scan_unevictable_node
,
3404 write_scan_unevictable_node
);
3406 int scan_unevictable_register_node(struct node
*node
)
3408 return device_create_file(&node
->dev
, &dev_attr_scan_unevictable_pages
);
3411 void scan_unevictable_unregister_node(struct node
*node
)
3413 device_remove_file(&node
->dev
, &dev_attr_scan_unevictable_pages
);